Memory-Aware Agent Harnesses for Anti-Money Laundering — Nacho Martinez, Oracle

Greetings. Hello everyone. Greetings. Hello everyone. Thank you for coming to Thank you for coming to Thank you for coming to this session. This will be a this session. This will be a this session. This will be a session on agent session on agent session on agent tools. We do tools. We do tools. We do n't have much time. The session n't have much time. The session n't have much time. The session will only last an will only last an will only last an hour. So I hour. So I hour. So I hope we will have time hope we will have time hope we will have time to discuss to discuss to discuss many interesting many interesting many interesting things for you. Essentially, this things for you. Essentially, this things for you. Essentially, this session will consist session will consist session will consist of me giving a of me giving a of me giving a 10-15 minute talk about the topics 10-15 minute talk about the topics 10-15 minute talk about the topics we will explore and we will explore and we will explore and address in the address in the address in the workshop. You have workshop. You have workshop. You have instructions for instructions for instructions for the workshop via QR code the workshop via QR code the workshop via QR code or on the website. If you or on the website. If you or on the website. If you don't want don't want don't want to listen to me, or you're to listen to me, or you're to listen to me, or you're bored, you can bored, you can bored, you can start working on the start working on the start working on the workshop yourself workshop yourself while I'm talking. But while I'm talking. But while I'm talking. But in general, here's what in general, here's what in general, here's what happens on the happens on the happens on the website. You simply website. You simply website. You simply register through register through register through your GitHub account and your GitHub account and your GitHub account and automatically automatically automatically receive an invitation receive an invitation receive an invitation to the repository. to the repository. The workshop is hosted on The workshop is hosted on CodeSpaces, so you can CodeSpaces, so you can CodeSpaces, so you can simply create a simply create a simply create a CodeSpace environment there. You CodeSpace environment there. You will automatically receive an invitation will automatically receive an invitation to access the to access the to access the workshop itself. You workshop itself. You workshop itself. You just need to just need to just need to create a CodeSpace there, create a CodeSpace there, create a CodeSpace there, open GitHub CodeSpaces, and open GitHub CodeSpaces, and open GitHub CodeSpaces, and you're ready to go you're ready to go you're ready to go and start your workshop.

and start your workshop. So, I'll display the So, I'll display the QR code again for those who QR code again for those who QR code again for those who want want want to use it. And I to use it. And I to use it. And I will begin. OK. will begin. OK. will begin. OK. Perfectly. If you have any Perfectly. If you have any Perfectly. If you have any questions, questions, questions, my colleague Richmond is there my colleague Richmond is there my colleague Richmond is there . It . It . It will fit will fit will fit every table. If every table. If every table. If you have a question, you have a question, you have a question, just raise your just raise your just raise your hand and he will hand and he will hand and he will help you. Yes. That's it help you. Yes. That's it . Also regarding the website: . Also regarding the website: . Also regarding the website: you can go to you can go to you can go to workshopwaitingroom.com. workshopwaitingroom.com. workshopwaitingroom.com. I paid too I paid too I paid too much money for this much money for this much money for this domain, so domain, so domain, so use it. use it. OK. So, who am I? OK. So, who am I? My name is Nacho, Nacho My name is Nacho, Nacho My name is Nacho, Nacho Martinez. I am an AI Martinez. I am an AI Martinez. I am an AI developer advocate and a data developer advocate and a data scientist at heart. And scientist at heart. And scientist at heart. And if you're interested in if you're interested in if you're interested in learning more about me learning more about me learning more about me , just , just , just find me on YouTube. find me on YouTube. find me on YouTube. I have podcasts, I have podcasts, I have podcasts, workshops, workshops, workshops, whatever. I won't whatever. I won't whatever. I won't tell you much tell you much tell you much about myself. Just one thing: about myself. Just one thing: about myself. Just one thing: we recently we recently we recently released a course released a course released a course with Andrew Yin and with Andrew Yin and with Andrew Yin and DeepLearning.AI on DeepLearning.AI on DeepLearning.AI on building adaptive building adaptive building adaptive AI agents, which is one AI agents, which is one AI agents, which is one of the components of the components of the components we'll we'll we'll talk about today. And if talk about today. And if talk about today. And if you're interested in what you're interested in what you're interested in what you can do with a you can do with a you can do with a model to model to model to improve it, improve it, improve it, adapt it, and adapt it, and adapt it, and train it over time, train it over time, train it over time, just check out just check out just check out this course. This is a this course. This is a this course. This is a free course, an

free course, an free course, an hour-long course that hour-long course that hour-long course that you can find on you can find on you can find on Google. So, what will we Google. So, what will we Google. So, what will we look at today? look at today? look at today? As I said, we do As I said, we do As I said, we do n't have much time, so I'll n't have much time, so I'll n't have much time, so I'll try to be try to be try to be concise. But here's what concise. But here's what concise. But here's what we're going to talk about. we're going to talk about. we're going to talk about. First of all, what is an First of all, what is an First of all, what is an agent stack? What agent stack? What agent stack? What components are included in the components are included in the components are included in the agent stack? What agent stack? What agent stack? What different forms can an different forms can an AI application take? What is an AI application take? What is an artificial artificial artificial intelligence agent? And what is an intelligence agent? And what is an intelligence agent? And what is an agent agent agent harness? What is the main harness? What is the main harness? What is the main motivation for creating motivation for creating motivation for creating or working with or working with or working with agent bindings agent bindings agent bindings rather than with AI rather than with AI agents themselves or agents themselves or agents themselves or LLM LLM models directly? And, you know, models directly? And, you know, models directly? And, you know, all the subcomponents of all the subcomponents of all the subcomponents of agent binding, like agent binding, like agent binding, like memory, memory, memory, context engineering, context engineering, context engineering, memory design, memory design, memory design, and all that stuff. and all that stuff. Since I was Since I was told that there would be told that there would be told that there would be many people from the many people from the many people from the financial sector, financial sector, financial sector, for example, many for example, many for example, many Wall Wall Street employees, I adapted Street employees, I adapted Street employees, I adapted this workshop to the this workshop to the this workshop to the banking topic. banking topic. So the idea is that the So the idea is that the basic basic basic premise of this premise of this premise of this workshop is this. We workshop is this. We workshop is this. We now often hear that, for now often hear that, for now often hear that, for example, OpenAI example, OpenAI example, OpenAI is launching 10,000 or 5,000 is launching 10,000 or 5,000 is launching 10,000 or 5,000 agents to agents to agents to solve a certain solve a certain solve a certain problem. 10,000 agents problem. 10,000 agents problem. 10,000 agents to solve a to solve a to solve a mathematical problem mathematical problem mathematical problem that could not be that could not be that could not be solved for 100 or 200 solved for 100 or 200 solved for 100 or 200 years. The thing is that these years. The thing is that these years. The thing is that these agents are autonomous.

agents are autonomous. They work without They work without human intervention until they human intervention until they human intervention until they reach reach reach a conclusion or an a conclusion or an a conclusion or an answer, right? answer, right? answer, right? These agents do not These agents do not These agents do not have the ability to have the ability to have the ability to do this on their own. Why? Because the do this on their own. Why? Because the do this on their own. Why? Because the context window context window context window is limited. is limited. is limited. Memory components are something that a Memory components are something that a Memory components are something that a large language model large language model large language model does not have naturally. does not have naturally. does not have naturally. So we have to implement all these things So we have to implement all these things ourselves ourselves to make agents to make agents to make agents capable of capable of capable of solving these kinds of solving these kinds of solving these kinds of problems autonomously within 10 problems autonomously within 10 problems autonomously within 10 or 100 hours, right? or 100 hours, right? The idea is that many The idea is that many people focus on people focus on people focus on advanced models, advanced models, advanced models, but my opinion is this: but my opinion is this: but my opinion is this: we don't need to we don't need to we don't need to focus focus focus on them. We can on them. We can on them. We can use them, but use them, but use them, but let's focus let's focus let's focus on the binding—it's the on the binding—it's the on the binding—it's the interesting part that interesting part that interesting part that can be customized can be customized can be customized however you want. In this however you want. In this however you want. In this application, the idea is application, the idea is application, the idea is that it is a that it is a that it is a medium-sized bank that medium-sized bank that medium-sized bank that occasionally occasionally occasionally experiences experiences experiences suspicious transactions, suspicious transactions, suspicious transactions, which fall into which fall into which fall into five types. five types.

For example, if a million dollars is deposited into For example, if a million dollars is deposited into your account your account , and then , and then over the course of three days, over the course of three days, $50,000 is automatically and magically withdrawn from an $50,000 is automatically and magically withdrawn from an ATM—that ATM—that ATM—that looks suspicious, does looks suspicious, does looks suspicious, does n't it? Because why n't it? Because why n't it? Because why invest a million invest a million invest a million dollars and then dollars and then dollars and then just just just distribute it like that? So distribute it like that? So distribute it like that? So the idea is that we have a the idea is that we have a the idea is that we have a medium-sized bank that has medium-sized bank that has medium-sized bank that has a lot of a lot of a lot of transactions, and these transactions, and these transactions, and these transactions need to transactions need to transactions need to be processed somehow be processed somehow be processed somehow autonomously or autonomously or autonomously or at least flagged at least flagged at least flagged so that the bank employee so that the bank employee so that the bank employee spends less time spends less time spends less time reviewing them, reviewing them, reviewing them, you know? So you know? So you know? So let's let's let's get into some theory. get into some theory. get into some theory. What is an agent stack, What is an agent stack, What is an agent stack, right? When we right? When we right? When we communicate with an AI communicate with an AI agent, we're usually agent, we're usually agent, we're usually dealing with the dealing with the dealing with the application layer, right? Er application layer, right? Er , something we , something we , something we interact with: interact with: interact with: an interface, a chatbot, a an interface, a chatbot, a an interface, a chatbot, a mobile app, mobile app, mobile app, etc. Then we have the etc. Then we have the etc. Then we have the model itself, which is the model itself, which is the model itself, which is the core of the reasoning, and we, core of the reasoning, and we, core of the reasoning, and we, you know, have no you know, have no you know, have no influence on its decisions.

influence on its decisions. Uh, and we also have Uh, and we also have the infrastructure to the infrastructure to the infrastructure to support all of this, and the support all of this, and the support all of this, and the computing computing computing power. But power. But power. But today we today we today we will focus on the will focus on the data part. Data is very important because Data is very important because everything we can everything we can everything we can manipulate and work with manipulate and work with is right here. is right here. We We We have less control over other parts, have less control over other parts, have less control over other parts, right? Especially if right? Especially if right? Especially if we use a we use a we use a third-party model that third-party model that third-party model that we don't own. So the we don't own. So the we don't own. So the data layer is what data layer is what data layer is what we're going to we're going to we're going to focus on focus on focus on today, right? First of today, right? First of today, right? First of all, we have the all, we have the all, we have the agent cycle, and this is probably agent cycle, and this is probably agent cycle, and this is probably one of the most important one of the most important one of the most important parts of an AI agent. Uh, the parts of an AI agent. Uh, the parts of an AI agent. Uh, the agent loop agent loop agent loop is responsible for is responsible for is responsible for managing the entire managing the entire managing the entire process, and it process, and it process, and it can also be can also be can also be customized. And customized. And customized. And inside the data, you inside the data, you inside the data, you also have a gateway and an MCP so also have a gateway and an MCP so also have a gateway and an MCP so you can you can you can connect to connect to connect to third-party external third-party external third-party external data sources. You data sources. You data sources. You also have a also have a also have a memory level, which we'll memory level, which we'll memory level, which we'll talk about later. And talk about later. And , in fact, some , in fact, some contextual contextual engineering components that engineering components that engineering components that allow you to allow you to allow you to automatically automatically automatically summarize, summarize, summarize, compress, and compress, and compress, and reduce the working reduce the working reduce the working context window context window context window of the model. And finally, of the model. And finally, of the model. And finally, some tools and some tools and some tools and skills that allow skills that allow skills that allow us to do us to do us to do interesting and clever interesting and clever interesting and clever things to limit things to limit things to limit the size of the context the size of the context the size of the context window. This is one of the window. This is one of the window. This is one of the main ideas of this main ideas of this main ideas of this workshop. Typically, workshop. Typically, workshop. Typically, the applications the applications the applications we interact with are we interact with are we interact with are a combination of these a combination of these a combination of these four elements,

four elements, four elements, right? LLM chatbots, which are right? LLM chatbots, which are right? LLM chatbots, which are very simple. Many very simple. Many very simple. Many people simply people simply people simply create an AI create an AI application, application, application, create a chatbot create a chatbot create a chatbot so that you can communicate with it so that you can communicate with it so that you can communicate with it . You . You . You also have RAG applications, also have RAG applications, also have RAG applications, which is essentially an which is essentially an which is essentially an enhanced vector enhanced vector enhanced vector search. And there are also search. And there are also search. And there are also LLM-based workflows and LLM-based workflows and LLM-based workflows and AI agents. And what AI agents. And what most of us usually use, by most of us usually use, by the way, raise your hands, the way, raise your hands, the way, raise your hands, who doesn't use AI who doesn't use AI agents like Claude Code agents like Claude Code agents like Claude Code and so on? Of course. So, and so on? Of course. So, and so on? Of course. So, Claude Code, Codex, etc.—all of Claude Code, Codex, etc.—all of Claude Code, Codex, etc.—all of these systems are a these systems are a these systems are a combination of the third combination of the third combination of the third and the fourth, that is, and the fourth, that is, and the fourth, that is, you know, a combination of you know, a combination of LLM-based workflows that LLM-based workflows that provide provide provide automation and AI automation and AI agents that agents that agents that provide provide provide autonomy. So, autonomy. So, autonomy. So, it's a combination of it's a combination of it's a combination of automation and automation and automation and autonomy. And such autonomy. And such autonomy. And such systems are called systems are called systems are called agent-based. The best agent-based. The best agent-based. The best systems, like Claude Code, systems, like Claude Code, systems, like Claude Code, Codex, Pi, Open Claude, Hermes Agent, Codex, Pi, Open Claude, Hermes Agent, Codex, Pi, Open Claude, Hermes Agent, etc., are all a etc., are all a etc., are all a combination of these two combination of these two combination of these two types. And most of the types. And most of the types. And most of the engineers I've engineers I've engineers I've talked to, and this is talked to, and this is talked to, and this is important, important, important, use these use these use these systems every day, systems every day, systems every day, right? But we right? But we right? But we stop, or a stop, or a stop, or a lot of people, lot of people, lot of people, actually not many actually not many actually not many people think about people think about people think about these issues when they these issues when they these issues when they build their AI build their AI applications. So applications. So applications. So you might have an AI you might have an AI application, but you're not application, but you're not application, but you're not using it that way using it that way using it that way . And that's what . And that's what . And that's what we're going to we're going to we're going to focus on today, because I focus on today, because I focus on today, because I think that's the best think that's the best think that's the best way to consume AI way to consume AI applications—not just applications—not just applications—not just through chatbots, not through chatbots, not through chatbots, not just through RAG just through RAG applications, but through a applications, but through a applications, but through a combination of combination of combination of automation and automation and automation and autonomy. So,

autonomy. So, autonomy. So, what is an AI agent for what is an AI agent for what is an AI agent for those who don't know? It is a those who don't know? It is a those who don't know? It is a computational computational computational entity that works entity that works entity that works on LLM, on on LLM, on on LLM, on reasoning, similar to reasoning, similar to reasoning, similar to us humans, we have our us humans, we have our us humans, we have our brains. It is supplemented with a brains. It is supplemented with a brains. It is supplemented with a database for database for database for memory, just like memory, just like memory, just like the brain. It is extended the brain. It is extended the brain. It is extended with with with action tools. action tools. action tools. That is, the equivalent That is, the equivalent That is, the equivalent for a human would be that for a human would be that for a human would be that we have limbs and we have limbs and we have limbs and we can we can we can interact with interact with interact with the world, just as an AI the world, just as an AI agent can agent can agent can interact with an interact with an interact with an operating system operating system operating system through the tools through the tools through the tools we provide it with. And we provide it with. And we provide it with. And it is also based it is also based it is also based on input data. on input data. on input data. For example, it would be For example, it would be For example, it would be like our human like our human like our human senses, eyes, senses, eyes, senses, eyes, etc., which allow etc., which allow etc., which allow us to perceive us to perceive us to perceive information. And information. And information. And context, like our context, like our context, like our attention, is also attention, is also attention, is also limited, as are limited, as are limited, as are our agents. We our agents. We our agents. We can't watch seven can't watch seven can't watch seven movies at the same time. We movies at the same time. We movies at the same time. We won't understand any of the won't understand any of the won't understand any of the films, we won't films, we won't films, we won't be able to focus be able to focus be able to focus on any of them. So the on any of them. So the on any of them. So the context window for context window for context window for humans is similar to AI humans is similar to AI agents. And we agents. And we agents. And we try to internalize try to internalize try to internalize or make some of or make some of or make some of these concepts closer these concepts closer these concepts closer to our human to our human to our human experience. And that's what experience. And that's what experience. And that's what we'll be we'll be we'll be working on today.

working on today. So, the part So, the part responsible for responsible for responsible for reasoning reasoning reasoning is provided by Frontier is provided by Frontier is provided by Frontier Labs. And we don't have Labs. And we don't have Labs. And we don't have much much much control over that. And then there are control over that. And then there are control over that. And then there are things we things we things we have a little more have a little more have a little more control over, namely control over, namely control over, namely memory, memory, memory, tool use, and tool use, and tool use, and perception, right? And this perception, right? And this perception, right? And this is the control system itself is the control system itself is the control system itself . The framework is . The framework is essentially the entire essentially the entire essentially the entire foundation foundation foundation needed to needed to needed to build these three build these three build these three components: components: tool use, memory, tool use, memory, and perception. And we won't and perception. And we won't focus too much on the focus too much on the model, because, as I model, because, as I model, because, as I said, we don't have said, we don't have said, we don't have much power to much power to much power to influence it, influence it, influence it, right? So, the framework right? So, the framework right? So, the framework is actually one of the is actually one of the is actually one of the most interesting topics in the most interesting topics in the most interesting topics in the field of AI. I think field of AI. I think field of AI. I think this is the most interesting topic in this is the most interesting topic in this is the most interesting topic in AI today, as AI today, as AI today, as these frameworks essentially these frameworks essentially these frameworks essentially drive everything an drive everything an drive everything an AI agent can do. You AI agent can do. You AI agent can do. You change change change the configuration, and the agent the configuration, and the agent —the same advanced —the same advanced —the same advanced LLM model you LLM model you LLM model you use—becomes use—becomes use—becomes better or worse.

better or worse. better or worse. Sometimes, you know, it was Sometimes, you know, it was Sometimes, you know, it was proven that by simply proven that by simply proven that by simply changing the framework, the changing the framework, the changing the framework, the same model same model same model worked better or worked better or worked better or worse on specific worse on specific worse on specific tasks. So it's really tasks. So it's really tasks. So it's really important to important to important to focus on the focus on the focus on the frameworks. And few frameworks. And few frameworks. And few people do that. And people do that. And people do that. And as you said, as you said, as you said, there are, you know, seven there are, you know, seven there are, you know, seven or eight different or eight different or eight different components in the framework, if components in the framework, if components in the framework, if you also consider you also consider you also consider continuous learning continuous learning . So, there are a lot of . So, there are a lot of . So, there are a lot of decisions you decisions you decisions you need to make, need to make, need to make, right? You have a model, right? You have a model, right? You have a model, you have a repository, i.e. you have a repository, i.e. you have a repository, i.e. where do you put the where do you put the where do you put the data? You have some data? You have some data? You have some coding coding coding and retrieval components, like and retrieval components, like and retrieval components, like how do I calculate how do I calculate how do I calculate the distance between two the distance between two the distance between two vectors to vectors to vectors to get the get the get the data correctly? What exactly do I data correctly? What exactly do I data correctly? What exactly do I keep in my keep in my keep in my memory? Should we memory? Should we memory? Should we think of it as think of it as think of it as short-term memory short-term memory short-term memory plus long-term memory, plus long-term memory, plus long-term memory, similar to how it similar to how it similar to how it works in humans? Should I works in humans? Should I works in humans? Should I consider consider consider different bays or different bays or different bays or different types of memory? different types of memory? Also the semantic Also the semantic level, which is level, which is level, which is really interesting really interesting really interesting because we as humans because we as humans because we as humans are here and are here and are here and we have a certain we have a certain we have a certain context that we've context that we've context that we've never talked about to never talked about to never talked about to each other, right? But each other, right? But each other, right? But we share it.

we share it. we share it. For example, I am For example, I am For example, I am the speaker and you the speaker and you the speaker and you are listening. You are not are listening. You are not are listening. You are not singing now. Why aren't you singing now. Why aren't you singing now. Why aren't you singing? Because it's singing? Because it's semantically inappropriate in the context semantically inappropriate in the context we're in right now we're in right now we're in right now , , , right? So, the model itself right? So, the model itself right? So, the model itself also has access to a also has access to a also has access to a semantic level, semantic level, semantic level, which explains a little bit of which explains a little bit of which explains a little bit of everything that was not everything that was not everything that was not told to the model directly, told to the model directly, told to the model directly, but is implied. but is implied. Also, the agent cycle is Also, the agent cycle is everything that controls it and everything that controls it and everything that controls it and makes the agent makes the agent makes the agent work work work automatically and automatically and automatically and autonomously over and over again autonomously over and over again autonomously over and over again . And this is one of the . And this is one of the . And this is one of the most important parts most important parts . And also . And also . And also designing the designing the designing the context. We will context. We will context. We will focus on focus on focus on each of them, but I each of them, but I each of them, but I will not will not will not go into go into go into detail, as we will detail, as we will detail, as we will do that in the do that in the do that in the seminar. So, for seminar. So, for seminar. So, for reference, there was a big reference, there was a big reference, there was a big debate about files debate about files debate about files vs. databases, vs. databases, vs. databases, for example, many people for example, many people for example, many people said we said we said we need to need to need to use use use files for everything. It files for everything. It files for everything. It 's like you just put 's like you just put 's like you just put everything in a file, and then, as if by magic everything in a file, and then, as if by magic , a month later , a month later the agent becomes better.

the agent becomes better. the agent becomes better. Other people are Other people are database maximalists database maximalists database maximalists , they want to , they want to , they want to cram everything into the cram everything into the cram everything into the database, right? So, this database, right? So, this database, right? So, this dilemma is still ongoing, dilemma is still ongoing, dilemma is still ongoing, and I would say, well, and I would say, well, and I would say, well, let me give you an let me give you an let me give you an example. Can example. Can example. Can someone tell me, in someone tell me, in someone tell me, in this example that this example that this example that you see on the screen, you see on the screen, you see on the screen, can someone tell me can someone tell me what workaround the what workaround the agents use to solve agents use to solve this particular this particular this particular problem that you problem that you problem that you see on the screen— see on the screen— multiple agents multiple agents multiple agents writing to the writing to the writing to the same file? How does it same file? How does it same file? How does it work for work for work for agents nowadays? Any ideas? agents nowadays? Any ideas? Work trees, Work trees, for sure. Good. So, for sure. Good. So, for sure. Good. So, working trees are the working trees are the working trees are the current solution, current solution, current solution, right? And this is a kind of right? And this is a kind of right? And this is a kind of workaround workaround workaround because files have an because files have an because files have an internal problem: internal problem: internal problem: if multiple records if multiple records if multiple records go into the go into the go into the same file, same file, same file, there are there are there are consistency issues. Like, consistency issues. Like, consistency issues. Like, which entry was the first which entry was the first ? If someone reads, what should ? If someone reads, what should ? If someone reads, what should I do with this I do with this I do with this person who person who person who read the file read the file read the file before I changed it before I changed it , and so on. A lot of , and so on. A lot of , and so on. A lot of problems, right? So problems, right? So problems, right? So for certain things, I for certain things, I for certain things, I will definitely tell you that will definitely tell you that will definitely tell you that files are great.

files are great. files are great. Files are great Files are great Files are great for for for storing large storing large storing large amounts of amounts of amounts of unstructured unstructured unstructured data. Databases, on data. Databases, on data. Databases, on the other hand, the other hand, the other hand, are also used for are also used for are also used for other types of tasks, other types of tasks, other types of tasks, but my solution is that but my solution is that files are suitable for certain cases. For other files are suitable for certain cases. For other cases, they cases, they cases, they work with work with work with databases. So we databases. So we databases. So we actually need to actually need to actually need to rethink some rethink some rethink some of these things that are of these things that are of these things that are happening, right? happening, right? And the solution or And the solution or workaround with workaround with workaround with working trees working trees working trees was already solved by was already solved by was already solved by databases 30 years databases 30 years databases 30 years ago. And we ago. And we ago. And we are trying to reinvent are trying to reinvent are trying to reinvent the wheel, and we the wheel, and we the wheel, and we don't really don't really don't really need to need to need to invent anything, because there invent anything, because there invent anything, because there is already a ready-made is already a ready-made is already a ready-made solution for this. For example, solution for this. For example, solution for this. For example, if you try to if you try to if you try to write to a write to a write to a file simultaneously by multiple file simultaneously by multiple file simultaneously by multiple agents. If you agents. If you agents. If you had transactional had transactional had transactional consistency or ACID consistency or ACID consistency, meaning consistency, meaning consistency, meaning atomic, consistent, atomic, consistent, atomic, consistent, isolated, and isolated, and isolated, and durable writes to a durable writes to a durable writes to a file, you wouldn't have file, you wouldn't have file, you wouldn't have this problem, right? this problem, right? this problem, right? So, files are good for So, files are good for So, files are good for some things, and some things, and some things, and databases are good for others. We databases are good for others. We databases are good for others. We don't have to don't have to don't have to choose just choose just choose just one thing, right? So in the one thing, right? So in the one thing, right? So in the agent system, agent system, agent system, just take this just take this just take this as food for thought.

as food for thought. Another component is Another component is all the all the all the memory components, right? And memory components, right? And memory components, right? And that encompasses all the that encompasses all the that encompasses all the mechanics and systems mechanics and systems mechanics and systems that are built that are built that are built around an AI agent around an AI agent around an AI agent to make it to make it to make it use use use information, and not just information, and not just information, and not just use it, but use it, but use it, but be able to be able to be able to refine and refine and refine and reproduce it in a timely manner when reproduce it in a timely manner when reproduce it in a timely manner when needed, right needed, right ? And agent memory—you, ? And agent memory—you, ? And agent memory—you, as engineers, as engineers, as engineers, choose what exactly you choose what exactly you choose what exactly you will consider in the will consider in the will consider in the memory level of your memory level of your memory level of your agent. Some people agent. Some people agent. Some people like the division into like the division into like the division into short-term and short-term and short-term and long-term long-term long-term memory. And in this memory. And in this memory. And in this case, it's part case, it's part case, it's part two. We have certain two. We have certain two. We have certain components of components of components of short-term short-term short-term memory. For example, memory. For example, memory. For example, if we are interacting with if we are interacting with if we are interacting with Claude or Codex, we have Claude or Codex, we have Claude or Codex, we have short-term short-term short-term memory in the sense memory in the sense memory in the sense that the context that the context that the context window itself is window itself is window itself is short-term short-term short-term memory. Everything that memory. Everything that memory. Everything that disappears after...it disappears after...it disappears after...it disappears after the conversation, disappears after the conversation, disappears after the conversation, right? Most right? Most right? Most things in the context things in the context things in the context window simply window simply window simply disappear. Then you disappear. Then you disappear. Then you start a new start a new start a new conversation and it conversation and it conversation and it starts from scratch.

starts from scratch. starts from scratch. This is This is This is short-term short-term short-term memory. It is very memory. It is very memory. It is very limited. We limited. We limited. We can't fit everything into it can't fit everything into it can't fit everything into it . And then we . And then we . And then we also have also have also have long-term long-term long-term memory. And within memory. And within memory. And within long-term long-term long-term memory, we also memory, we also memory, we also offer some offer some offer some categories. For example, categories. For example, categories. For example, episodic memory episodic memory is all the events that is all the events that is all the events that have happened, all the episodes of have happened, all the episodes of have happened, all the episodes of our conversations, all the our conversations, all the our conversations, all the tool logs, tool logs, tool logs, previous conversations, previous conversations, previous conversations, summaries of previous summaries of previous summaries of previous conversations you had with conversations you had with conversations you had with the agent, etc. All of this the agent, etc. All of this the agent, etc. All of this falls into this falls into this falls into this category of category of category of episodic episodic episodic long-term long-term long-term memory. You'll have memory. You'll have memory. You'll have procedures or procedures or procedures or procedural memory— procedural memory— those are all the things like those are all the things like those are all the things like skills, skills, skills, workflows, which are workflows, which are workflows, which are sequences of sequences of sequences of skills that skills that skills that are performed, and so on. are performed, and so on. are performed, and so on. All of this goes into All of this goes into All of this goes into procedural memory. procedural memory. procedural memory. And we also have And we also have And we also have semantic memory, semantic memory, semantic memory, which is everything else, as I which is everything else, as I which is everything else, as I said. The agent said. The agent said. The agent remembers who you are, what remembers who you are, what remembers who you are, what your preferences are, what your preferences are, what your preferences are, what your, you know, your, you know, your, you know, favorite framework is, favorite framework is, favorite framework is, etc. All of these things etc. All of these things etc. All of these things go into go into go into semantic memory. semantic memory. So the problem, the So the problem, the main question here is: main question here is: main question here is: why can't we why can't we why can't we put everything in put everything in put everything in short-term short-term short-term memory? Because it is memory? Because it is memory? Because it is limited. And not limited. And not limited. And not only is it limited, but only is it limited, but only is it limited, but in fact, the in fact, the in fact, the larger the attention matrix becomes larger the attention matrix becomes larger the attention matrix becomes (and the (and the (and the attention matrix is, in essence, a attention matrix is, in essence, a attention matrix is, in essence, a context window), context window), context window), the worse it works the worse it works . And this happens with . And this happens with . And this happens with every model. There are every model. There are every model. There are many studies many studies many studies that have proven that if you that have proven that if you that have proven that if you start with zero start with zero start with zero tokens in the tokens in the tokens in the context window, context window, context window, the problem is solved the problem is solved the problem is solved much more accurately than much more accurately than much more accurately than if you already have 256,000 if you already have 256,000 if you already have 256,000 tokens. And now

tokens. And now tokens. And now the limit is one the limit is one the limit is one million. We have a million. We have a million. We have a million thanks to the million thanks to the million thanks to the KV cache. All of these things should KV cache. All of these things should KV cache. All of these things should be taken into account, and just be taken into account, and just be taken into account, and just know: even if the know: even if the know: even if the stated limit stated limit stated limit stops at 256K stops at 256K stops at 256K or 1 million, if or 1 million, if or 1 million, if you keep you keep you keep increasing the volume, increasing the volume, increasing the volume, performance becomes performance becomes performance becomes terrible. So, the terrible. So, the terrible. So, the smaller the context smaller the context smaller the context window, the better it window, the better it window, the better it will solve the problem. And will solve the problem. And will solve the problem. And that applies to that applies to any type of problem any type of problem imaginable in humanity, imaginable in humanity, right? And there are two types of right? And there are two types of right? And there are two types of approaches that approaches that approaches that solve this. The first solve this. The first is is is context summarization, where you context summarization, where you context summarization, where you simply ask, "Hey, simply ask, "Hey, simply ask, "Hey, generate me a summary, generate me a summary, generate me a summary, a quick summary of a quick summary of a quick summary of everything that happened in the everything that happened in the everything that happened in the context window," context window," context window," and then you get and then you get and then you get a summary from which you can a summary from which you can a summary from which you can start the next start the next start the next conversation. This is perfectly conversation. This is perfectly conversation. This is perfectly acceptable. Another acceptable. Another acceptable. Another approach is compression, approach is compression, approach is compression, and it's somewhat similar, but and it's somewhat similar, but and it's somewhat similar, but with compression you with compression you with compression you just put just put just put parts of the conversation in a parts of the conversation in a parts of the conversation in a database or file, database or file, database or file, assign them assign them assign them an ID, and if the an ID, and if the an ID, and if the AI ​​agent AI ​​agent AI ​​agent needs it later, needs it later, needs it later, it will use that ID.

it will use that ID. it will use that ID. It simply extracts It simply extracts It simply extracts data by data by data by ID with an ID with an ID with an offset. It's a bit like offset. It's a bit like offset. It's a bit like reading from a file, reading from a file, reading from a file, right? These are just some right? These are just some right? These are just some of the challenges that of the challenges that of the challenges that arise during arise during arise during context engineering. context engineering. And then we have the And then we have the agent loop, which is the driver agent loop, which is the driver agent loop, which is the driver or the machine that controls or the machine that controls or the machine that controls all the components that all the components that all the components that we talked about. we talked about. we talked about. We have a part of the We have a part of the We have a part of the context collection that takes context collection that takes context collection that takes data from skills, data from skills, data from skills, tools, tools, tools, previous conversations, previous conversations, previous conversations, files, databases— files, databases— all of this is dynamically all of this is dynamically all of this is dynamically entered into the entered into the entered into the context window. context window. Then, Then, reasoning occurs, an reasoning occurs, an API call is made to OpenAI or whatever API call is made to OpenAI or whatever model model model you are working with, and you are working with, and you are working with, and it is determined what it is determined what it is determined what needs to be done. And needs to be done. And needs to be done. And it happens, it happens, it happens, you know, automatically and you know, automatically and you know, automatically and endlessly, if you endlessly, if you endlessly, if you wish. In my wish. In my wish. In my case, this case, this case, this agent loop that we agent loop that we agent loop that we prepared for the prepared for the prepared for the workshop has a workshop has a workshop has a maximum maximum maximum time limit time limit time limit because I don't want to because I don't want to because I don't want to run out of run out of run out of OpenAI credits or any other ones OpenAI credits or any other ones OpenAI credits or any other ones that I have. But yes, the that I have. But yes, the that I have. But yes, the maximum maximum maximum number of iterations in number of iterations in number of iterations in our case is our case is eight, and a maximum of 60 eight, and a maximum of 60 eight, and a maximum of 60 seconds of reasoning seconds of reasoning seconds of reasoning for each iteration, for each iteration, for each iteration, understood? But this is understood? But this is understood? But this is completely completely completely customizable. As I customizable. As I customizable. As I said, as engineers we said, as engineers we said, as engineers we have to decide what have to decide what have to decide what exactly goes into the exactly goes into the exactly goes into the agent loop, right? As well as the agent loop, right? As well as the agent loop, right? As well as the " " " toolbar" and " toolbar" and " skillbar" patterns skillbar" patterns skillbar" patterns we offer for we offer for we offer for storing hundreds of storing hundreds of storing hundreds of tools and tools and tools and skills. I don't know about skills. I don't know about skills. I don't know about you, but I already have you, but I already have you, but I already have about about about a hundred skills installed. I don't

a hundred skills installed. I don't a hundred skills installed. I don't put a hundred put a hundred put a hundred skills in the context window at once skills in the context window at once because each skill because each skill takes up about takes up about 3,000–5,000 tokens. So, this would 3,000–5,000 tokens. So, this would 3,000–5,000 tokens. So, this would instantly exceed the instantly exceed the instantly exceed the stated limit in a stated limit in a stated limit in a matter of seconds. matter of seconds. matter of seconds. I won't have time I won't have time I won't have time to actually to actually to actually talk to the agent. talk to the agent. talk to the agent. So, how to fix this? So, how to fix this? Well, there are different ways, Well, there are different ways, but we suggest but we suggest but we suggest vectorizing, that is, vectorizing, that is, vectorizing, that is, creating vector creating vector creating vector embeddings of embeddings of embeddings of tools and tools and tools and skills. And thanks to skills. And thanks to skills. And thanks to these vector these vector these vector embeddings, once the embeddings, once the embeddings, once the agent loop... Okay, agent loop... Okay, agent loop... Okay, let me let me let me go back. go back. go back. When the agent loop wants to When the agent loop wants to When the agent loop wants to get useful get useful get useful skills or skills or skills or tools, you tools, you tools, you do a do a do a vector similarity search vector similarity search vector similarity search on this vector on this vector on this vector repository that you have. repository that you have. repository that you have. You choose the You choose the You choose the most appropriate most appropriate most appropriate skill or skill or skill or tool to tool to tool to solve the task. solve the task. solve the task. You solve You solve You solve the problem and then the problem and then the problem and then upload upload upload it again. And that's one it again. And that's one it again. And that's one way you way you way you keep the keep the context window relevant in context window relevant in that sense. And, you know that sense. And, you know , using , using , using skills and skills and skills and tools—that's a tools—that's a tools—that's a very, very common very, very common very, very common pattern. For example, pattern. For example, pattern. For example, everyone uses everyone uses everyone uses tools. Like the tools. Like the tools. Like the bash command or changing bash command or changing bash command or changing directory, these are all directory, these are all bash-specific tools, bash-specific tools, right? So we are right? So we are right? So we are constantly constantly constantly using using using tools and tools and tools and skills. In some skills. In some skills. In some cases, that's over 50% of the cases, that's over 50% of the context window used, context window used, right? So I just right? So I just right? So I just want you to be want you to be want you to be aware of the aware of the aware of the huge amount of huge amount of huge amount of context window context window context window that that that skills and tools actually take up skills and tools actually take up skills and tools actually take up . Um, we . Um, we . Um, we also have an idea also have an idea also have an idea called called called skill distillation,

skill distillation, skill distillation, where you take where you take where you take previous conversations, previous conversations, previous conversations, performances, and, you know, performances, and, you know, performances, and, you know, real conversations that real conversations that real conversations that you've had with an AI agent, and you've had with an AI agent, and you've had with an AI agent, and try to create try to create try to create something something something meaningful out of it. How to meaningful out of it. How to meaningful out of it. How to do it? Well, there are do it? Well, there are do it? Well, there are a few a few a few ways to do this, but one ways to do this, but one ways to do this, but one of the easiest is to of the easiest is to of the easiest is to simply create a simply create a simply create a skill.md file and put it skill.md file and put it skill.md file and put it in the same in the same in the same skills folder where you had skills folder where you had skills folder where you had everything before. And this is one everything before. And this is one everything before. And this is one of the ways that Agent of the ways that Agent of the ways that Agent Hermes, one of the Hermes, one of the Hermes, one of the innovators in this innovators in this innovators in this regard, created regard, created regard, created skills that are particularly skills that are particularly skills that are particularly well suited to well suited to well suited to solving solving solving specific problems. specific problems. specific problems. Let's say you're Let's say you're Let's say you're working on a working on a working on a frontend. Well, if frontend. Well, if frontend. Well, if you work on you work on you work on the frontend, you can the frontend, you can the frontend, you can create a skill that has create a skill that has worked really well for you with worked really well for you with all your all your all your preferences. Hey, preferences. Hey, preferences. Hey, I like this I like this I like this style of colors and such. style of colors and such. style of colors and such. And the next time And the next time And the next time you you you work on work on work on the frontend, since the frontend, since the frontend, since you created vector you created vector you created vector embeddings for it, embeddings for it, embeddings for it, the agent will automatically the agent will automatically the agent will automatically choose the choose the choose the most appropriate most appropriate most appropriate skill. So if you skill. So if you skill. So if you just say, "Hey, just say, "Hey, just say, "Hey, I like that I like that I like that color style that color style that color style that was in that other was in that other was in that other skill," it will skill," it will skill," it will automatically understand automatically understand that you mean the that you mean the that you mean the new skill because new skill because new skill because you used the word you used the word you used the word "color style," "color style," "color style," for example, right?

for example, right? So, So, skill distillation is one skill distillation is one skill distillation is one technique that can be technique that can be technique that can be used to used to used to improve improve improve model learning over time. And it's the same model learning over time. And it's the same model learning over time. And it's the same . This is related . This is related . This is related to what I was talking about: to what I was talking about: to what I was talking about: you have dynamic you have dynamic you have dynamic skill discovery, skill discovery, skill discovery, meaning the agent, during meaning the agent, during meaning the agent, during context formation, context formation, context formation, selects the right selects the right selects the right skills and skills and skills and tools when tools when tools when needed. And you know, needed. And you know, needed. And you know, some people ask, some people ask, some people ask, how can a frozen model how can a frozen model how can a frozen model get better? get better? get better? By "frozen" I By "frozen" I By "frozen" I mean that we mean that we mean that we have no have no have no control over the control over the control over the boundary models, boundary models, boundary models, right? But we have right? But we have right? But we have control over the entire control over the entire control over the entire data layer, right? data layer, right? Some say, "Hey, Some say, "Hey, let's change the let's change the let's change the model weights." Others say, model weights." Others say, model weights." Others say, "Let's build an "Let's build an "Let's build an embedding model embedding model embedding model that works better for that works better for that works better for this particular this particular this particular task, and then I'll task, and then I'll task, and then I'll just worry about just worry about just worry about getting better getting better getting better embeddings." So there embeddings." So there embeddings." So there are many are many are many ways to do this.

ways to do this. But the idea is that But the idea is that it is a core it is a core it is a core component of component of component of continuous continuous continuous learning. And learning. And learning. And continuous learning continuous learning is just one of the things you is just one of the things you can do can do can do to improve the model. to improve the model. And there are three categories, And there are three categories, right? You are either right? You are either right? You are either working with working with working with parametric parametric parametric memory. It memory. It memory. It is called is called is called parametric parametric parametric because it changes because it changes because it changes the parameters of the the parameters of the the parameters of the neural network weights. You neural network weights. You neural network weights. You also have also have also have representational representational representational memory that works with memory that works with memory that works with embeddings and so on. And embeddings and so on. And embeddings and so on. And you also have you also have you also have nonparametric nonparametric nonparametric memory, which we've memory, which we've memory, which we've talked about so far today talked about so far today talked about so far today . But . But . But if you're interested in if you're interested in if you're interested in any of these three, any of these three, any of these three, we've focused on the we've focused on the we've focused on the third one. You also third one. You also third one. You also have two others, but I wo have two others, but I wo have two others, but I wo n't go into them n't go into them n't go into them . . . There are many techniques, There are many techniques, There are many techniques, such as creating a such as creating a such as creating a low-rank low-rank low-rank adaptation, adaptation, adaptation, model training, model training, model training, creating embeddings, creating embeddings, creating embeddings, etc. You'll etc. You'll etc. You'll find all of this in the course find all of this in the course find all of this in the course I mentioned I mentioned I mentioned earlier, which earlier, which earlier, which focuses on these focuses on these focuses on these three categories. So three categories. So three categories. So if you are interested, if you are interested, if you are interested, take this course or take this course or take this course or chat with us chat with us chat with us after the session. We'll after the session. We'll after the session. We'll be at the event be at the event be at the event throughout the entire time throughout the entire time throughout the entire time after the session, right? And one after the session, right? And one after the session, right? And one last thing I want to last thing I want to last thing I want to mention: when you mention: when you mention: when you call the chat call the chat complement API to any complement API to any complement API to any OpenAI model or other OpenAI model or other OpenAI model or other model you model you model you use, use, use, there is a built-in there is a built-in there is a built-in hierarchy of instructions.

hierarchy of instructions. What do I mean by What do I mean by instruction hierarchy instruction hierarchy ? There is a different weight for the ? There is a different weight for the ? There is a different weight for the things that are given to things that are given to things that are given to the model. For example, what the model. For example, what the model. For example, what we we we interact with is the interact with is the interact with is the user, right? user, right? user, right? Request. In this sense, Request. In this sense, Request. In this sense, this is the third most this is the third most this is the third most important thing important thing important thing we can convey to the we can convey to the we can convey to the model. The most important one model. The most important one is the system prompt is the system prompt , right? Everyone knows , right? Everyone knows , right? Everyone knows about this, the concept of a about this, the concept of a about this, the concept of a system prompt. system prompt. Um, the system prompt is good, but it can't is good, but it can't is good, but it can't accommodate everything, accommodate everything, accommodate everything, otherwise there would be no otherwise there would be no otherwise there would be no hierarchy inside. You hierarchy inside. You hierarchy inside. You actually need to actually need to actually need to put some put some put some things in the system things in the system things in the system prompt, some in the prompt, some in the prompt, some in the user prompt, user prompt, user prompt, and all the tools, and all the tools, and all the tools, skills, and things that are skills, and things that are skills, and things that are provided have their own provided have their own provided have their own hierarchy. And that's a good thing hierarchy. And that's a good thing hierarchy. And that's a good thing because it actually because it actually because it actually allows us or the allows us or the allows us or the model model model to focus more on to focus more on to focus more on some things than some things than some things than others. And we will also others. And we will also others. And we will also consider this during the consider this during the consider this during the workshop. So, for workshop. So, for workshop. So, for reference, this is the interface of reference, this is the interface of reference, this is the interface of our workshop. So our workshop. So our workshop. So if you're already if you're already if you're already running it in running it in running it in CodeSpaces, you'll see that CodeSpaces, you'll see that CodeSpaces, you'll see that they're deployed to you they're deployed to you . I'll walk you through all of this . I'll walk you through all of this . I'll walk you through all of this .

. But this is essentially the But this is essentially the way I way I way I imagine a person imagine a person imagine a person working in AML, like in a working in AML, like in a fraud detection system at a bank, fraud detection system at a bank, would interact with an would interact with an would interact with an autonomous agent. The autonomous agent. The standalone agent standalone agent runs in the background runs in the background runs in the background 24/7. He is not 24/7. He is not 24/7. He is not sleeping. He doesn't sleeping. He doesn't sleeping. He doesn't need any of the need any of the need any of the human needs that human needs that human needs that we have, right? It is we have, right? It is we have, right? It is connected via an connected via an connected via an interface to the interface to the interface to the database and has all the database and has all the database and has all the functionality functionality we mentioned. But we mentioned. But beyond that, he beyond that, he beyond that, he autonomously makes autonomously makes autonomously makes decisions for the company, decisions for the company, decisions for the company, even when, you know, even when, you know, even when, you know, we're offline. He doesn't we're offline. He doesn't we're offline. He doesn't ruin people's lives by ruin people's lives by ruin people's lives by just saying, "Hey, just saying, "Hey, just saying, "Hey, this guy is this guy is this guy is suspicious. "Let's suspicious. "Let's suspicious. "Let's just block the just block the just block the whole bank and everything, whole bank and everything, whole bank and everything, you know, because that would you know, because that would you know, because that would ruin a ruin a ruin a person's life." Instead, person's life." Instead, person's life." Instead, it makes these it makes these it makes these decisions and simply decisions and simply decisions and simply records them so that the person records them so that the person records them so that the person can check them can check them can check them later. But it later. But it later. But it will save a lot of will save a lot of will save a lot of time during, you know, the time during, you know, the time during, you know, the discovery process discovery process discovery process and so on. So by the end of and so on. So by the end of and so on. So by the end of the workshop, if you the workshop, if you the workshop, if you complete it, you complete it, you complete it, you will be able to create will be able to create will be able to create an agent that an agent that an agent that connects to a connects to a connects to a database, works like database, works like database, works like in any company, in any company, in any company, and performs autonomous and performs autonomous and performs autonomous analysis and analysis and analysis and decision-making, which a decision-making, which a decision-making, which a human can then human can then human can then review and review and review and process. So, those are process. So, those are process. So, those are all the components all the components all the components we talked about we talked about we talked about regarding our regarding our regarding our interface. And without

interface. And without interface. And without further ado, I'll further ado, I'll further ado, I'll bring up the QR code again or bring up the QR code again or bring up the QR code again or workshopwaitingroom.com. And if workshopwaitingroom.com. And if workshopwaitingroom.com. And if you want you want you want to check out to check out to check out our other our other our other workshops, you can workshops, you can workshops, you can simply check out simply check out simply check out this repository on this repository on this repository on GitHub. We put in a lot of GitHub. We put in a lot of GitHub. We put in a lot of effort and effort and effort and progress in everything progress in everything progress in everything we add there. we add there. From workshops and From workshops and sessions to notebooks sessions to notebooks sessions to notebooks to get you to get you to get you started on a variety of started on a variety of started on a variety of topics like agent topics like agent topics like agent systems and more. So systems and more. So systems and more. So just check it out. just check it out. Let me Let me jump into Firefox and jump into Firefox and jump into Firefox and show you how show you how show you how this application works. So, this application works. So, this application works. So, first of all, first of all, first of all, let me say let me say let me say to those who didn't get to those who didn't get to those who didn't get an invite: if you an invite: if you an invite: if you go to go to go to workshopwaitingroom.com, you workshopwaitingroom.com, you workshopwaitingroom.com, you will get an invitation will get an invitation will get an invitation via the GitHub API for via the GitHub API for via the GitHub API for your user your user your user to this repository. to this repository. So if you just So if you just go here and go here and go here and then open Codespaces in GitHub then open Codespaces in GitHub then open Codespaces in GitHub , it will ask , it will ask , it will ask you to create a Codespace, and you to create a Codespace, and you to create a Codespace, and of course I'm paying for it, of course I'm paying for it, of course I'm paying for it, so don't worry.

so don't worry. Be careful with the Be careful with the equipment you equipment you equipment you deploy because deploy because deploy because it is expensive. The idea it is expensive. The idea it is expensive. The idea is that you is that you is that you get this get this get this deployment with a couple of deployment with a couple of deployment with a couple of laptops, I'll show laptops, I'll show laptops, I'll show you the workshop itself, and you the workshop itself, and you the workshop itself, and then after a few then after a few then after a few minutes I'll step away and we'll minutes I'll step away and we'll minutes I'll step away and we'll go through go through go through the questions, if any, the questions, if any, the questions, if any, and solve them together. and solve them together. But it automatically But it automatically deploys this deploys this deploys this application, which application, which application, which essentially tracks all essentially tracks all essentially tracks all actions, and this is, of course, actions, and this is, of course, actions, and this is, of course, fake data. Don't fake data. Don't fake data. Don't worry, I'm worry, I'm worry, I'm not exposing anyone here, but not exposing anyone here, but not exposing anyone here, but it essentially generates it essentially generates it essentially generates new transactions. new transactions. new transactions. Some of them are normal Some of them are normal . Some of them are . Some of them are . Some of them are problematic. So, an problematic. So, an problematic. So, an autonomous agent, autonomous agent, autonomous agent, which is a Grok model, which is a Grok model, which is a Grok model, running on OCI. It running on OCI. It running on OCI. It essentially makes essentially makes essentially makes autonomous decisions, and autonomous decisions, and autonomous decisions, and about some of them it says, " about some of them it says, " Hey, that's suspicious. Hey, that's suspicious. Hey, that's suspicious. Let's add that to the Let's add that to the Let's add that to the list." And then, you list." And then, you list." And then, you also have also have also have the option the option the option to visualize the to visualize the agent context window here. You agent context window here. You can also view can also view can also view all the data already in the all the data already in the all the data already in the database here.

database here. For example, all For example, all customers, etc. And customers, etc. And customers, etc. And also an autonomous also an autonomous also an autonomous way in which the agent way in which the agent way in which the agent will analyze all of will analyze all of will analyze all of this in real time, this in real time, this in real time, right? So it right? So it right? So it will use the will use the will use the tool calls tool calls that we provide, that are that we provide, that are that we provide, that are available in our available in our available in our toolbox. toolbox. Yes, sorry, it's Yes, sorry, it's a bit small here. He a bit small here. He a bit small here. He will go through the will go through the will go through the evidence. It will evidence. It will evidence. It will search for knowledge in the search for knowledge in the search for knowledge in the knowledge spaces that knowledge spaces that knowledge spaces that we created we created we created using vector using vector using vector repositories. It will make repositories. It will make repositories. It will make many many many tool calls and tool calls and tool calls and eventually automatically eventually automatically eventually automatically review all review all review all activity and activity and activity and suspicious suspicious suspicious activity reports. And here's activity reports. And here's activity reports. And here's just an interesting just an interesting just an interesting way to visualize the way to visualize the way to visualize the 30 most recent 30 most recent 30 most recent events that happened at the events that happened at the events that happened at the bank. But bank. But bank. But the seminar itself is this the seminar itself is this the seminar itself is this notebook, a notebook, a notebook, a student notebook student notebook . ipynb. And I want you to . ipynb. And I want you to . ipynb. And I want you to first just first just first just look at the different look at the different look at the different components that we components that we components that we have there, because we're have there, because we're have there, because we're implementing them right here implementing them right here . And I'll go through the . And I'll go through the . And I'll go through the first two. There are a few first two. There are a few first two. There are a few things you things you things you need to do at the need to do at the need to do at the workshop, just to workshop, just to workshop, just to confirm that you confirm that you confirm that you know what you are doing.

know what you are doing. So, first of all, So, first of all, let me let me let me show you the first show you the first show you the first part of the workshop, part of the workshop, part of the workshop, okay? You have this okay? You have this okay? You have this Grok 4.3 model Grok 4.3 model Grok 4.3 model running on Oracle Cloud running on Oracle Cloud . And let's, of course, . And let's, of course, . And let's, of course, connect to the connect to the connect to the database. And after... database. And after... Excuse me, let Excuse me, let Excuse me, let me restart me restart me restart the kernel. Let me the kernel. Let me run this run this cell. cell. cell. Restart. But Restart. But Restart. But yeah, let us know yeah, let us know yeah, let us know if you have any if you have any if you have any questions or questions or questions or problems problems problems getting into the code space, and getting into the code space, and getting into the code space, and then I'll come over and then I'll come over and then I'll come over and do it with you and do it with you and do it with you and answer any answer any answer any questions. By the way, are there questions. By the way, are there questions. By the way, are there any any any questions now? Good. Yes? questions now? Good. Yes?

Repository? Repository? Yes. Yes. Yes. Yes. You can find Yes. You can find Yes. You can find it at workshopwaitingroom.com, it at workshopwaitingroom.com, it at workshopwaitingroom.com, and let me and let me and let me just put the just put the just put the QR code here. Here's this QR QR code here. Here's this QR code if you want to code if you want to code if you want to scan it. And scan it. And scan it. And this this this will automatically give you access to will automatically give you access to will automatically give you access to it via the GitHub API. it via the GitHub API. Okay, let me Okay, let me rephrase the rephrase the rephrase the question for people question for people question for people who are on YouTube. So he who are on YouTube. So he who are on YouTube. So he asks what asks what asks what happens to happens to happens to episodic memory episodic memory episodic memory and all these and all these and all these components of memory. components of memory. components of memory. Is she confusing this with Is she confusing this with Is she confusing this with previous previous previous conversations with the agent? conversations with the agent? Well, it depends on Well, it depends on how how how much weight you much weight you much weight you give it, and that's why I give it, and that's why I give it, and that's why I also mentioned the also mentioned the also mentioned the hierarchy of instructions. hierarchy of instructions. hierarchy of instructions. You might want to You might want to You might want to put this in, if put this in, if put this in, if it's a really, really important it's a really, really important it's a really, really important thing and you really thing and you really thing and you really want it to come want it to come want it to come from from from somewhere, you somewhere, you somewhere, you can actually can actually design the memory design the memory layer architecture yourself layer architecture yourself layer architecture yourself , right? So, you , right? So, you , right? So, you could, for example, in could, for example, in could, for example, in cases where cases where cases where episodic memory is episodic memory is episodic memory is very important and you very important and you very important and you really want to really want to really want to use use use previous conversations, you previous conversations, you previous conversations, you could vectorize could vectorize could vectorize that instead of just that instead of just that instead of just storing it as storing it as storing it as plain text, right?

plain text, right? plain text, right? A lot depends on the A lot depends on the A lot depends on the use case use case , I would say. In this , I would say. In this , I would say. In this case, there is no case, there is no case, there is no conflict conflict conflict because we have a clear because we have a clear because we have a clear allocation of allocation of allocation of memory components in the architecture memory components in the architecture . Yes, skills can . Yes, skills can . Yes, skills can be added to a system be added to a system be added to a system query. Usually they query. Usually they ...Oh, sorry. ...Oh, sorry. ...Oh, sorry. They usually have another...Yes, They usually have another...Yes, They usually have another...Yes, in the in the in the instruction hierarchy they instruction hierarchy they instruction hierarchy they have the lowest weight of the have the lowest weight of the have the lowest weight of the four. Well, if you four. Well, if you four. Well, if you think about it, it's think about it, it's think about it, it's , um, usually right , um, usually right , um, usually right here. So it has the here. So it has the here. So it has the least weight. I would least weight. I would least weight. I would call it a plug-and-play call it a plug-and-play call it a plug-and-play place where you can place where you can place where you can take anything. And we're take anything. And we're take anything. And we're experimenting with experimenting with experimenting with this, actually me and I, this, actually me and I, this, actually me and I, because we want to create a because we want to create a because we want to create a way to connect way to connect way to connect to external or to external or to external or third-party third-party third-party skill centers and skill centers and skill centers and tools. tools. For example, a collection of For example, a collection of MCPs that you can MCPs that you can MCPs that you can give your agent give your agent give your agent access to by access to by precomputing vector precomputing vector representations ( representations ( embeddings) for embeddings) for embeddings) for all of it. And when all of it. And when all of it. And when you have these you have these you have these pre-computed pre-computed pre-computed embeddings, you embeddings, you embeddings, you can just say, "Hey, can just say, "Hey, can just say, "Hey, do a vector do a vector do a vector search across these search across these search across these extraneous extraneous extraneous skill sets or skill sets or skill sets or tools." But tools." But tools." But instead of instead of instead of thinking of it as thinking of it as thinking of it as included in the request—which, by the way, is included in the request—which, by the way, is included in the request—which, by the way, is of course of course of course included in every included in every included in every request to OpenAI or the ChatComplete API request to OpenAI or the ChatComplete API that you that you that you send to OpenAI or send to OpenAI or send to OpenAI or any other any other any other model—it's added model—it's added model—it's added at the end. And usually this at the end. And usually this at the end. And usually this is about 80% is about 80% of the volume. That is, 80% of the of the volume. That is, 80% of the entire request is entire request is entire request is actually tools

actually tools actually tools and skills. But what's and skills. But what's and skills. But what's happening, or why are happening, or why are happening, or why are n't you being charged n't you being charged n't you being charged five times more five times more five times more than you are now? This is because than you are now? This is because than you are now? This is because OpenAI and all these systems OpenAI and all these systems OpenAI and all these systems have a semantic have a semantic have a semantic cache. And they cache. And they cache. And they keep...if 80% of keep...if 80% of the query is the same as the the query is the same as the the query is the same as the previous one, they don't previous one, they don't previous one, they don't calculate it again. calculate it again. It just goes into the It just goes into the semantic cache, and it semantic cache, and it semantic cache, and it allows us all allows us all allows us all to use to use to use Frontier Labs models in a Frontier Labs models in a Frontier Labs models in a cost- cost- cost- effective way. effective way. And of course, it's still And of course, it's still very expensive, but I very expensive, but I very expensive, but I want you to understand: want you to understand: want you to understand: it would be four or it would be four or five times more expensive five times more expensive five times more expensive if we didn't have if we didn't have if we didn't have this semantic this semantic this semantic cache where cache where cache where these these these tools and tools and tools and skills go. Because in skills go. Because in skills go. Because in response, response, response, remember that when remember that when remember that when you interact with an you interact with an you interact with an advanced AI model, advanced AI model, advanced AI model, you are essentially doing you are essentially doing you are essentially doing everything through an HTTP request. And everything through an HTTP request. And everything through an HTTP request. And all this, for example, 10 all this, for example, 10 all this, for example, 10 thousand or 5 thousand thousand or 5 thousand thousand or 5 thousand system system system prompt tokens. You are an AI prompt tokens. You are an AI assistant, and you have to assistant, and you have to assistant, and you have to help me help me help me solve this, that, and that.

solve this, that, and that. solve this, that, and that. You have instructions from You have instructions from You have instructions from the developer for a the developer for a the developer for a specific specific specific application, a application, a application, a user request—what user request—what user request—what we type in the chat. we type in the chat. we type in the chat. Which is meager, Which is meager, Which is meager, maybe 100 or 1000 maybe 100 or 1000 maybe 100 or 1000 tokens. And then tokens. And then tokens. And then the results of the the results of the the results of the tools and the tools and the tools and the extracted documents, extracted documents, extracted documents, which make up 80% of which make up 80% of the query, that is, from 35 to the query, that is, from 35 to the query, that is, from 35 to 100 thousand tokens in 100 thousand tokens in 100 thousand tokens in each query. Does this each query. Does this each query. Does this answer your answer your answer your question? question? question? Last question. Last question. Last question. Yes. Good. So, his Yes. Good. So, his Yes. Good. So, his question was about question was about how do you manage a how do you manage a how do you manage a budget? And since there's budget? And since there's so so much in the skill set and tools, how do you much in the skill set and tools, how do you much in the skill set and tools, how do you even make it even make it even make it work, huh? Um, work, huh? Um, work, huh? Um, we have a clear division. we have a clear division. We perceive the context window We perceive the context window as an n- as an n- dimensional matrix that dimensional matrix that dimensional matrix that starts at zero starts at zero starts at zero but is never but is never but is never zero. It's closer zero. It's closer zero. It's closer to 10,000. I read to 10,000. I read to 10,000. I read an article an article an article comparing the comparing the comparing the performance of each performance of each performance of each framework based on framework based on framework based on how much they add how much they add how much they add to the first system to the first system to the first system prompt and the first request prompt and the first request prompt and the first request to the completion API. At the to the completion API. At the to the completion API. At the very beginning of Pi, very beginning of Pi, very beginning of Pi, you know Pi, this framework?

you know Pi, this framework? you know Pi, this framework? This is the most This is the most This is the most minimalistic and minimalistic and minimalistic and efficient framework efficient framework efficient framework because it includes because it includes because it includes about 6,000 tokens about 6,000 tokens about 6,000 tokens at the beginning of each at the beginning of each at the beginning of each request. But if request. But if request. But if you take other frameworks, they you take other frameworks, they you take other frameworks, they usually have 30-40 usually have 30-40 usually have 30-40 thousand tokens. Um, how do thousand tokens. Um, how do thousand tokens. Um, how do we...we have a clear we...we have a clear we...we have a clear division. Context division. Context division. Context window, we just window, we just window, we just let everything let everything let everything revolve around the revolve around the revolve around the agent loop, right? So agent loop, right? So agent loop, right? So let me let me let me go into the go into the go into the agent cycle to show agent cycle to show agent cycle to show you one thing. The you one thing. The you one thing. The agent loop will control agent loop will control agent loop will control everything it everything it everything it connects to, right? Um, connects to, right? Um, connects to, right? Um, so here, for example, is the so here, for example, is the so here, for example, is the agent cycle. Oh, I'm at the very agent cycle. Oh, I'm at the very agent cycle. Oh, I'm at the very bottom. bottom. bottom. Wait. Uh, just Wait. Uh, just Wait. Uh, just as a reference, uh, as a reference, uh, as a reference, uh, agent cycle So, here, right? agent cycle So, here, right? Uh, the context assembly, Uh, the context assembly, when you work with it when you work with it when you work with it , uh, you're , uh, you're , uh, you're essentially building, and you essentially building, and you essentially building, and you can walk through it in a can walk through it in a can walk through it in a laptop. This laptop. This laptop. This is explained in is explained in is explained in the notebook, but, uh, the notebook, but, uh, the notebook, but, uh, context collection—this is the context collection—this is the context collection—this is the part where, first of all, part where, first of all, part where, first of all, we have a maximum we have a maximum we have a maximum number of certain number of certain number of certain uh tools or uh tools or uh tools or skills to skills to skills to obtain. The obtain. The default number seems to be three. Uh, so every three. Uh, so every three. Uh, so every time a time a time a new request comes in, it new request comes in, it new request comes in, it determines, and the determines, and the determines, and the model itself does that. It model itself does that. It model itself does that. It determines exactly what determines exactly what determines exactly what needs to be done.

needs to be done. needs to be done. Then she Then she Then she tries to put together tries to put together tries to put together that context with the that context with the that context with the things she things she things she needs to needs to needs to execute, right? And one execute, right? And one execute, right? And one of those things is of those things is of those things is getting the three getting the three getting the three most relevant most relevant most relevant tools, the three tools, the three tools, the three most relevant most relevant most relevant skills. And we don't skills. And we don't skills. And we don't use a use a use a reorderer. reorderer. But what you could also But what you could also do is, do is, do is, for example, at the end of a for example, at the end of a for example, at the end of a similarity search, similarity search, similarity search, add an add an add an embedding model plus a embedding model plus a embedding model plus a reranker reranker reranker that takes these three and that takes these three and that takes these three and tells you, "Of these tells you, "Of these tells you, "Of these three, this is the three, this is the three, this is the most important one." "And you most important one." "And you most important one." "And you could actually could actually could actually get better get better get better refinement because you refinement because you refinement because you might not want to might not want to might not want to load a load a load a total of three skills total of three skills total of three skills for each task for each task ." The task can ." The task can ." The task can only be solved by one, only be solved by one, only be solved by one, right? So the idea is that right? So the idea is that right? So the idea is that you can you can you can use use any strategy you any strategy you any strategy you want. Just uh want. Just uh want. Just uh stick to it and uh stick to it and uh stick to it and uh be consistent be consistent be consistent until the very end. So until the very end. So until the very end. So the idea in this the idea in this the idea in this particular workshop particular workshop particular workshop is that we is that we is that we perform a certain perform a certain perform a certain number of vector number of vector number of vector searches in two searches in two searches in two vector repositories: vector repositories: vector repositories: one containing one containing one containing tools and the other containing tools and the other containing skills. K is equal to skills. K is equal to skills. K is equal to three, and from these three three, and from these three three, and from these three you can extract the you can extract the you can extract the most relevant one most relevant one most relevant one using a using a using a reranker. And reranker. And reranker. And then, of course, it then, of course, it then, of course, it distributes distributes distributes the tools and the tools and the tools and forms a context forms a context forms a context window, but that's window, but that's window, but that's entirely the decision and entirely the decision and entirely the decision and work of the agent cycle.

work of the agent cycle. Yes. So, what is the Yes. So, what is the most important or most important or most important or interesting way interesting way interesting way to interact with the to interact with the to interact with the memory level? So memory level? So memory level? So the efficiency the efficiency the efficiency only depends on only depends on only depends on the number of entities the number of entities the number of entities you create, you create, you create, right? So my right? So my right? So my first recommendation would be to first recommendation would be to first recommendation would be to just not just not just not complicate things. Don't complicate things. Don't complicate things. Don't create too create too create too many abstractions many abstractions many abstractions or too many or too many or too many components for the components for the components for the memory layer. memory layer. memory layer. For example, try For example, try experimenting experimenting with just with just with just short-term and short-term and short-term and long-term long-term long-term memory first. And just memory first. And just memory first. And just have two structures, have two structures, have two structures, right? If you think, right? If you think, right? If you think, or if you have an agent or if you have an agent that decides that that decides that that decides that something from something from something from short-term short-term short-term memory or memory or memory or context is context is context is worth saving, you worth saving, you worth saving, you move it into move it into move it into long-term long-term long-term memory and just memory and just memory and just store it as store it as store it as plain text. This is the plain text. This is the plain text. This is the easiest thing to easiest thing to easiest thing to do, right? do, right? do, right? Once it appears Once it appears Once it appears as text somewhere in as text somewhere in as text somewhere in long-term long-term long-term memory or an object, you memory or an object, you memory or an object, you can take it, can take it, can take it, vectorize it, and vectorize it, and vectorize it, and then let the then let the then let the agent loop agent loop agent loop explore it in some explore it in some explore it in some way. For example, here way. For example, here way. For example, here you would have an you would have an you would have an extra "stuff" extra "stuff" extra "stuff" between the between the between the observation and observation and observation and reasoning stages that reasoning stages that reasoning stages that allows the agent allows the agent allows the agent to recall memories or to recall memories or to recall memories or something like that. You something like that. You something like that. You could call this, could call this, could call this, for example, " for example, " for example, " long-term long-term long-term memory recall." And this will force memory recall." And this will force memory recall." And this will force him to be him to be him to be introspective introspective introspective before before before solving the problem again.

solving the problem again. So, it will turn So, it will turn to your vector to your vector to your vector storage or some other storage or some other storage or some other place where this place where this place where this episode of episode of episode of long-term long-term long-term memory, recorded memory, recorded memory, recorded during a previous during a previous during a previous conversation, lies. And from that, you conversation, lies. And from that, you conversation, lies. And from that, you can let can let can let him decide, " him decide, " Okay, I'll let Okay, I'll let Okay, I'll let him pull this out him pull this out him pull this out if he finds it if he finds it if he finds it interesting," or if the interesting," or if the interesting," or if the vector similarity, vector similarity, vector similarity, that is, the distance between the that is, the distance between the that is, the distance between the query vector and what's query vector and what's in in in long-term long-term long-term memory, is memory, is memory, is small enough, then I'll add it to the small enough, then I'll add it to the small enough, then I'll add it to the context window. context window. If it's too If it's too big or not big or not big or not relevant enough, I'll just relevant enough, I'll just relevant enough, I'll just throw it away. So, this would throw it away. So, this would throw it away. So, this would be an additional be an additional be an additional abstraction between the abstraction between the observation and observation and reasoning part. There are ways in reasoning part. There are ways in which we're which we're misusing MCP here, misusing MCP here, but one of the benefits of but one of the benefits of but one of the benefits of using MCP is that it using MCP is that it using MCP is that it solves... Well, I would solves... Well, I would solves... Well, I would say that MCP is not say that MCP is not say that MCP is not very relevant here. But it very relevant here. But it very relevant here. But it depends on what depends on what depends on what you have in your database. you have in your database. you have in your database. So, for example, if So, for example, if So, for example, if you have a database and you have a database and you have a database and you interact with it you interact with it you interact with it like we do now, like we do now, like we do now, using MCP is using MCP is using MCP is interesting. One of the interesting. One of the interesting. One of the main questions main questions main questions I often ask people is: I often ask people is: I often ask people is: what exactly does MCP solve, what exactly does MCP solve, what exactly does MCP solve, especially in especially in especially in databases, since I am a databases, since I am a databases, since I am a database specialist, database specialist, database specialist, working at Oracle. So, working at Oracle. So, working at Oracle. So, one of the problems that one of the problems that one of the problems that can be solved is to can be solved is to can be solved is to allow AI agents allow AI agents allow AI agents to execute SQL queries to execute SQL queries to execute SQL queries without hallucinating and without hallucinating and without hallucinating and trying to trying to trying to invent a new SQL invent a new SQL query every time. This is one of the query every time. This is one of the query every time. This is one of the biggest problems biggest problems biggest problems AI agents have with

AI agents have with AI agents have with SQL. There are about 15 engines in SQL SQL. There are about 15 engines in SQL SQL. There are about 15 engines in SQL . Such as Postgres, . Such as Postgres, . Such as Postgres, Oracle, VSQL; there are many Oracle, VSQL; there are many Oracle, VSQL; there are many different queries that different queries that different queries that run on SQL, but run on SQL, but run on SQL, but they have they have they have specific hints specific hints specific hints or nuances that make or nuances that make or nuances that make them not always them not always them not always work well, work well, work well, you know? So the you know? So the you know? So the main problem that main problem that main problem that MCP solves in this MCP solves in this MCP solves in this case is: okay, I'll case is: okay, I'll case is: okay, I'll create SQL queries create SQL queries create SQL queries in advance and in advance and in advance and let agents let agents let agents call them. call them. So, these queries So, these queries are guaranteed to work are guaranteed to work , instead of the , instead of the , instead of the agent agent agent trying to create a SQL trying to create a SQL query dynamically each time. query dynamically each time. query dynamically each time. Right? Ugh. Yes, Right? Ugh. Yes, Right? Ugh. Yes, good question. good question. good question. Let me give you an Let me give you an Let me give you an example. example. The most important thing would be The most important thing would be the speed and content of the speed and content of the speed and content of the sentences, as well as the actual the sentences, as well as the actual the sentences, as well as the actual meaning of the skill, meaning of the skill, meaning of the skill, instead of instead of instead of storing it as text storing it as text . And if you were to store . And if you were to store . And if you were to store it simply as text in a it simply as text in a it simply as text in a SQL file, it would be SQL file, it would be SQL file, it would be possible...sorry, possible...sorry, possible...sorry, in a file on the file in a file on the file in a file on the file system, right? In an system, right? In an system, right? In an operating system, operating system, operating system, you would have a multitude of you would have a multitude of you would have a multitude of ways in which an ways in which an ways in which an agent agent agent could perform could perform regular expression searches.

For example, all searching would For example, all searching would be based on either be based on either be based on either keywords or keywords or keywords or word searches in that word searches in that word searches in that skill, which is not bad. skill, which is not bad. But the bigger it But the bigger it gets, the harder it is for the gets, the harder it is for the gets, the harder it is for the agent. First, agent. First, agent. First, regular regular regular expression searches like grep are expression searches like grep are expression searches like grep are inefficient for inefficient for inefficient for huge files, and huge files, and huge files, and sometimes if the file sometimes if the file sometimes if the file system grows system grows system grows linearly, linearly, linearly, the performance the performance the performance of of of regular expression searches regular expression searches regular expression searches drops exponentially. drops exponentially. Well, not exponentially, Well, not exponentially, quadratically. Let's quadratically. Let's quadratically. Let's just say quadratically. just say quadratically. just say quadratically. But this is one of the problems that makes But this is one of the problems that makes But this is one of the problems that makes it it it not scale as well as not scale as well as file sizes increase. file sizes increase. Another point: if you Another point: if you Another point: if you pre- pre- pre- compute compute compute skill embeddings, you actually skill embeddings, you actually skill embeddings, you actually capture hidden meaning in them capture hidden meaning in them that we humans that we humans cannot comprehend. cannot comprehend. Instead of Instead of just describing it in just describing it in just describing it in text that we are text that we are text that we are comfortable reading, it comfortable reading, it comfortable reading, it also has also has also has 384 to 1000 384 to 1000 384 to 1000 dimensions pre-computed, and this dimensions pre-computed, and this dimensions pre-computed, and this embedding embedding embedding will actually hold the will actually hold the will actually hold the relationships between each relationships between each relationships between each token in the matrix, token in the matrix, token in the matrix, making them much more making them much more making them much more accessible, accessible, accessible, for example, for for example, for for example, for semantic search.

semantic search. And this brings you to And this brings you to using using using hybrid search, hybrid search, hybrid search, which is a combination of which is a combination of which is a combination of semantic search, semantic search, semantic search, i.e. embeddings, and i.e. embeddings, and i.e. embeddings, and keyword search keyword search keyword search . And this . And this . And this is called RRF—a type of is called RRF—a type of is called RRF—a type of hybrid search. hybrid search. It works much It works much better because it better because it better because it essentially uses essentially uses essentially uses both methods. And this...I both methods. And this...I both methods. And this...I can show you an can show you an can show you an example in the workshop example in the workshop example in the workshop because I created this because I created this because I created this visualization that visualization that visualization that explains exactly what I mean explains exactly what I mean explains exactly what I mean . With . With . With hybrid search, hybrid search, hybrid search, right? See? Yes. So, right? See? Yes. So, right? See? Yes. So, depending on...It's depending on...It's depending on...It's essentially a graph that essentially a graph that essentially a graph that combines combines combines keyword search keyword search keyword search and and and semantic search capabilities. semantic search capabilities. And that gap that And that gap that you see in the middle is everything you see in the middle is everything you see in the middle is everything that's displayed that's displayed that's displayed here in a hybrid here in a hybrid here in a hybrid search will be more search will be more search will be more relevant than relevant than relevant than the results of either the results of either method method method alone. So, alone. So, alone. So, to summarize: the to summarize: the to summarize: the more you can more you can more you can combine two different combine two different combine two different types of searches...Each types of searches...Each types of searches...Each has its own advantages has its own advantages , but if you combine , but if you combine , but if you combine them, you them, you them, you will essentially spend a little will essentially spend a little will essentially spend a little more more more computing computing computing resources, but resources, but resources, but the quality of the search results the quality of the search results will improve over time. I think will improve over time. I think about it from the perspective of AI about it from the perspective of AI agents. So I don't agents. So I don't agents. So I don't expect anyone in this expect anyone in this expect anyone in this room to start

room to start room to start building their own building their own building their own tools, tools, tools, because we already have because we already have because we already have great solutions on the great solutions on the great solutions on the market, right? We market, right? We market, right? We have Cloud Call, we have Codex. have Cloud Call, we have Codex. have Cloud Call, we have Codex. Everyone has their own Everyone has their own Everyone has their own preferences. By the way, preferences. By the way, preferences. By the way, raise your hands, who raise your hands, who raise your hands, who loves pie? OK. loves pie? OK. loves pie? OK. Smaller than I expected, Smaller than I expected, Smaller than I expected, but good. So, I would but good. So, I would but good. So, I would say that I'm looking at say that I'm looking at say that I'm looking at this from the perspective of AI this from the perspective of AI agents, and I want agents, and I want agents, and I want you to understand: you you to understand: you you to understand: you can create can create can create specialized specialized specialized tools for tools for tools for specific tasks. And specific tasks. And specific tasks. And creating such creating such creating such tools doesn't tools doesn't tools doesn't necessarily have to be necessarily have to be necessarily have to be as complicated as as complicated as as complicated as developing new Cloud developing new Cloud developing new Cloud Code. And then you Code. And then you Code. And then you run all your run all your run all your work in this work in this work in this special special special tool. It's better tool. It's better tool. It's better to simplify the problems to simplify the problems to simplify the problems you're trying you're trying you're trying to solve, and then to solve, and then to solve, and then create custom create custom create custom tools that tools that tools that connect to them. connect to them. So in this case I So in this case I think: if you are a think: if you are a think: if you are a consultant, then you are consultant, then you are consultant, then you are my ideal client my ideal client my ideal client for this workshop for this workshop for this workshop because you could because you could because you could create a custom create a custom create a custom tool that tool that tool that works great for a works great for a works great for a financial agent and financial agent and financial agent and has everything you need for has everything you need for has everything you need for financial tasks.

financial tasks. But if you have a different But if you have a different type of client, different type of client, different type of client, different data, or different data, or different data, or different constraints, or if constraints, or if constraints, or if you want to solve you want to solve you want to solve another another another Navier-Stokes equation or something Navier-Stokes equation or something Navier-Stokes equation or something like that, then there like that, then there like that, then there might be different might be different might be different constraints. For example, constraints. For example, constraints. For example, OpenAI tasks, OpenAI tasks, OpenAI tasks, mathematical problems, mathematical problems, mathematical problems, require require require automatic automatic automatic compression or compression or compression or automatic automatic automatic generalization of generalization of generalization of context, because these context, because these context, because these agents do not have a human to agents do not have a human to agents do not have a human to guide the guide the guide the process. So anything that process. So anything that addresses these addresses these issues should issues should include automatic context generalization by default. And context generalization by default. And it must also possess it must also possess it must also possess infinite or infinite or infinite or almost infinite almost infinite almost infinite patience. In this patience. In this patience. In this case, the case, the case, the agent's work cycle is simplified to a agent's work cycle is simplified to a agent's work cycle is simplified to a maximum of eight maximum of eight maximum of eight steps. All these steps. All these steps. All these considerations, right? considerations, right? considerations, right? It depends on the It depends on the It depends on the problem itself. Yes, I would problem itself. Yes, I would problem itself. Yes, I would start from scratch, because that way start from scratch, because that way start from scratch, because that way you can discard 90% of what you can discard 90% of what you don't you don't you don't need. So, I would need. So, I would need. So, I would definitely start like this. Everything definitely start like this. Everything definitely start like this. Everything is built around the is built around the is built around the chat chat chat completions API. Both Anthropic and OpenAI completions API. Both Anthropic and OpenAI completions API. Both Anthropic and OpenAI have created their own have created their own have created their own standards to standards to standards to counter this.

counter this. counter this. But everything revolves But everything revolves But everything revolves around this API. So around this API. So around this API. So if we view if we view if we view the way of communicating with the way of communicating with the way of communicating with the model simply as an the model simply as an the model simply as an additional plugin to the additional plugin to the additional plugin to the system, it becomes system, it becomes system, it becomes clearer that the clearer that the clearer that the smaller the number, the better the smaller the number, the better the smaller the number, the better the result. Does this result. Does this result. Does this answer your answer your answer your question? OK. question? OK. question? OK. So, let's do this. You So, let's do this. You So, let's do this. You start with a codebase start with a codebase start with a codebase in Codespace. Just in Codespace. Just in Codespace. Just set up the notebook set up the notebook set up the notebook yourself and I'll yourself and I'll yourself and I'll go around to everyone, and if go around to everyone, and if go around to everyone, and if anyone has any anyone has any anyone has any questions, we can questions, we can questions, we can chat less chat less chat less formally because I formally because I formally because I feel like I'm being feel like I'm being feel like I'm being watched. And you watched. And you watched. And you probably don't probably don't probably don't feel any pressure. So feel any pressure. So feel any pressure. So I'll take the pressure off you by I'll take the pressure off you by I'll take the pressure off you by just walking away. And just walking away. And just walking away. And if you are, yes, if you are, yes, if you are, yes, let's continue. We have let's continue. We have let's continue. We have little time. Oh my God. little time. Oh my God. Time flew by. But yes. Time flew by. But yes. Thank you. Thank you for Thank you. Thank you for Thank you. Thank you for the applause. Yes, thank you.

the applause. Yes, thank you. the applause. Yes, thank you. Thank you. I thought Thank you. I thought Thank you. I thought 15 or 20 minutes had passed, 15 or 20 minutes had passed, 15 or 20 minutes had passed, but it was a whole hour but it was a whole hour . So, uh, one thing. Uh . So, uh, one thing. Uh , so I'm going to , so I'm going to , so I'm going to go. I, I set go. I, I set go. I, I set everything up so that you can everything up so that you can everything up so that you can create a code create a code create a code space in a hotel, space in a hotel, space in a hotel, at home. You can, I'll at home. You can, I'll at home. You can, I'll pay for it myself. So pay for it myself. So pay for it myself. So you can, you know, you can, you know, you can, you know, use this use this use this workshop and play with workshop and play with workshop and play with it however you want. And I will it however you want. And I will it however you want. And I will pay for it. So don't pay for it. So don't pay for it. So don't worry about it. worry about it. worry about it. Just, uh, if you Just, uh, if you Just, uh, if you want want want to chat about to chat about to chat about some of the concepts behind some of the concepts behind some of the concepts behind agent work, agent work, agent work, just come by the just come by the just come by the Oracle booth. We will be Oracle booth. We will be Oracle booth. We will be here until Wednesday evening. here until Wednesday evening. Uh, thank you for coming Uh, thank you for coming to the session. Yes. OK.