Decoupling the Brain from the Hands: Claude Managed Agents (Part 1 of 4) — Anthropic

Thank you very much to everyone who Thank you very much to everyone who joined. I think joined. I think joined. I think we'll start, we'll start, we'll start, since we have a full since we have a full since we have a full hall. Hello everyone. hall. Hello everyone. hall. Hello everyone. My name is Priyanka. I am a My name is Priyanka. I am a My name is Priyanka. I am a technical technical technical employee at Anthropic. employee at Anthropic. employee at Anthropic. I lead the I lead the I lead the engineering team that engineering team that engineering team that built built built Claude Managed Agents, and I'm Claude Managed Agents, and I'm Claude Managed Agents, and I'm excited to be here to excited to be here to excited to be here to chat with chat with chat with you and help you you and help you you and help you launch your launch your launch your first Managed first Managed first Managed Agent. I've been at Anthropic for Agent. I've been at Anthropic for Agent. I've been at Anthropic for a little over two a little over two a little over two years, and for the last 12 years, and for the last 12 years, and for the last 12 months I've months I've months I've been focused on been focused on been focused on building out the agent building out the agent building out the agent infrastructure, and infrastructure, and infrastructure, and I hope I hope I hope to share to share to share our our our work with you. work with you. work with you. Congratulations. My name is Congratulations. My name is Congratulations. My name is Gabe Semack. I am a Gabe Semack. I am a Gabe Semack. I am a technical technical technical fellow building fellow building Claude managed agents on Claude managed agents on Priyanka's team. I've been at Anthropic for Priyanka's team. I've been at Anthropic for Priyanka's team. I've been at Anthropic for a little over a year and a little over a year and have been working on this have been working on this project almost all that time. project almost all that time. OK. So, here is our OK. So, here is our agenda. We'll agenda. We'll agenda. We'll start with a history of start with a history of start with a history of how we how we how we built and built and built and scaled scaled scaled managed agents, managed agents, managed agents, as well as some of the as well as some of the as well as some of the pitfalls and pitfalls and pitfalls and lessons we lessons we lessons we learned along the learned along the learned along the way. Then we way. Then we way. Then we will move on to the will move on to the will move on to the practical part of practical part of practical part of the seminar. In the first the seminar. In the first the seminar. In the first part, we will create a part, we will create a part, we will create a very simple agent very simple agent very simple agent using the using the using the Claude platform console. You Claude platform console. You Claude platform console. You will not write will not write will not write any code during any code during any code during this part. We'll this part. We'll this part. We'll just try to just try to just try to create our create our create our first agent first agent first agent using the prompts and using the prompts and using the prompts and test it.

test it. test it. Then I'll hand it over to Then I'll hand it over to Then I'll hand it over to Gabe to walk Gabe to walk Gabe to walk you through the practical you through the practical you through the practical part. We'll be part. We'll be part. We'll be writing code writing code writing code using Claude using Claude using Claude Coder, your Coder, your Coder, your favorite agent favorite agent favorite agent for programming, and for programming, and for programming, and then building a then building a then building a more complex more complex more complex financial agent, financial agent, financial agent, after which we'll have after which we'll have after which we'll have plenty of time plenty of time plenty of time at the end for at the end for at the end for questions and questions and questions and answers. We have a answers. We have a answers. We have a two-hour slot. I two-hour slot. I two-hour slot. I know it's know it's know it's divided into four divided into four divided into four parts, but it's parts, but it's parts, but it's essentially one long essentially one long essentially one long presentation. presentation. presentation. I hope it doesn't I hope it doesn't I hope it doesn't take all two hours. take all two hours. take all two hours. We want it to be We want it to be We want it to be quite interactive. quite interactive. quite interactive. So if you have any So if you have any So if you have any questions, questions, questions, raise your hand. raise your hand. raise your hand. We have a few people We have a few people We have a few people around who can around who can around who can come up to you and give you come up to you and give you come up to you and give you a microphone for a microphone for a microphone for a question, especially a question, especially a question, especially when we get to the when we get to the when we get to the practical part. practical part. Okay, let's Okay, let's get started. So, I'll get started. So, I'll get started. So, I'll start with the basics. start with the basics. start with the basics. Let's start with what is an Let's start with what is an Let's start with what is an agent or what agent or what agent or what makes an agent makes an agent makes an agent work? An agent is work? An agent is work? An agent is essentially a model in a essentially a model in a essentially a model in a loop with loop with loop with tools. You tools. You tools. You will see on the left will see on the left will see on the left a model that makes a model that makes a model that makes a decision, performs the a decision, performs the a decision, performs the next action, and next action, and next action, and calls tools calls tools calls tools in a loop until your in a loop until your in a loop until your task is task is task is completed. This is essentially the completed. This is essentially the completed. This is essentially the simplest simplest simplest definition of what an definition of what an definition of what an agent is. Yes.

agent is. Yes. Link to Link to the presentation? We the presentation? We the presentation? We can try... can try... can try... We can share We can share We can share it when it's finished. it when it's finished. it when it's finished. Yes, we will share Yes, we will share Yes, we will share it when it is finished. It it when it is finished. It it when it is finished. It will be difficult for us to provide will be difficult for us to provide will be difficult for us to provide this now, but we will be this now, but we will be this now, but we will be happy to answer happy to answer happy to answer questions and questions and questions and share share share the presentation the presentation the presentation at the end. Do at the end. Do at the end. Do you have difficulty seeing slides from the you have difficulty seeing slides from the you have difficulty seeing slides from the back rows? Good. back rows? Good. back rows? Good. Sorry about that. Um, Sorry about that. Um, Sorry about that. Um, okay. The next thing okay. The next thing okay. The next thing I wanted to talk about I wanted to talk about is what a “harness” is is what a “harness” is . Harness is . Harness is . Harness is usually a layer that usually a layer that usually a layer that wraps around the model and wraps around the model and wraps around the model and helps execute helps execute helps execute tool calls tool calls tool calls in a loop to in a loop to in a loop to achieve the achieve the achieve the result of your result of your result of your current task. current task. current task. It makes It makes It makes assumptions about assumptions about assumptions about what the model cannot what the model cannot what the model cannot do and tries do and tries do and tries to fill in those to fill in those to fill in those gaps. So, it gaps. So, it gaps. So, it works in a loop, works in a loop, works in a loop, making making making tool calls, tool calls, tool calls, managing context, managing context, managing context, and encapsulating anything and encapsulating anything and encapsulating anything the model can't the model can't the model can't do on its own. A do on its own. A do on its own. A good example is good example is good example is our Sonnet 3.5 model, our Sonnet 3.5 model, our Sonnet 3.5 model, which had what we which had what we which had what we called “ called “ contextual alerting contextual alerting .” So whenever .” So whenever .” So whenever Sonnet 3.5 sensed that Sonnet 3.5 sensed that Sonnet 3.5 sensed that you were approaching the you were approaching the you were approaching the context limit, context limit, context limit, it tried it tried it tried to complete to complete to complete the task very quickly. So when the task very quickly. So when the task very quickly. So when creating a wrapper creating a wrapper creating a wrapper around Sonnet 3.5, you around Sonnet 3.5, you around Sonnet 3.5, you added something like added something like context reset to get around context reset to get around this limitation. But with the this limitation. But with the this limitation. But with the next model, like next model, like Opus 3.5, there Opus 3.5, there Opus 3.5, there was no such limitation, and was no such limitation, and was no such limitation, and you didn't need to you didn't need to you didn't need to do such do such do such context resetting for it.

context resetting for it. context resetting for it. So when you build a So when you build a So when you build a wrapper, you want to wrapper, you want to wrapper, you want to encapsulate all these encapsulate all these encapsulate all these assumptions for assumptions for assumptions for different models and different models and different models and keep up with each keep up with each keep up with each new model, so new model, so new model, so building a good building a good building a good wrapper is very wrapper is very wrapper is very important, and that's what we important, and that's what we important, and that's what we tried to do with tried to do with tried to do with Cloud Managed Agents. And lastly Cloud Managed Agents. And lastly : you want your : you want your : you want your agent to have access to a agent to have access to a agent to have access to a sandbox or sandbox or sandbox or computer. Essentially, computer. Essentially, computer. Essentially, this is where he will this is where he will this is where he will run the code he run the code he run the code he writes to perform the writes to perform the writes to perform the task. Typically, you task. Typically, you task. Typically, you provide access to a provide access to a provide access to a cloud sandbox cloud sandbox cloud sandbox or container where or container where or container where it executes the it executes the it executes the written code. It written code. It written code. It can edit can edit can edit files, install files, install files, install dependencies, etc. Cloud dependencies, etc. Cloud dependencies, etc. Cloud Code is a great example of Code is a great example of Code is a great example of this. This is a wrapper this. This is a wrapper this. This is a wrapper running locally on running locally on running locally on your computer, your computer, your computer, accessing your accessing your accessing your files and executing them, files and executing them, files and executing them, updating them in the process. updating them in the process. updating them in the process. He also writes code He also writes code He also writes code that is unreliable. that is unreliable. that is unreliable. It can be It can be It can be buggy, so when buggy, so when buggy, so when deploying deploying deploying product agents product agents product agents you want to you want to you want to make sure the code make sure the code make sure the code they write they write they write is is is sandboxed and sandboxed and sandboxed and confined to confined to confined to the container. You also the container. You also the container. You also need to keep your need to keep your need to keep your sandbox sandbox sandbox isolated and your isolated and your isolated and your attack radius limited.

attack radius limited. attack radius limited. Therefore, you typically Therefore, you typically Therefore, you typically want to run want to run want to run each task in a each task in a each task in a separate container separate container separate container to avoid mixing data to avoid mixing data to avoid mixing data and increasing the and increasing the and increasing the impact radius for impact radius for impact radius for your tasks. Next, I'll your tasks. Next, I'll your tasks. Next, I'll tell you about the tell you about the tell you about the first architecture on first architecture on first architecture on which we actually which we actually which we actually built managed built managed built managed agents. It was agents. It was agents. It was last December last December last December when we delivered the first when we delivered the first when we delivered the first architecture to our architecture to our architecture to our early early early access customers, got access customers, got access customers, got a ton of feedback, and a ton of feedback, and a ton of feedback, and then moved on to the then moved on to the then moved on to the new architecture. So new architecture. So new architecture. So I'll tell you about I'll tell you about I'll tell you about some of the some of the some of the pitfalls and problems pitfalls and problems pitfalls and problems we faced in the we faced in the we faced in the first version. In the first version. In the first version. In the first architecture, first architecture, first architecture, we placed everything in a we placed everything in a we placed everything in a container. So, container. So, container. So, we had a session log we had a session log we had a session log with everything that with everything that with everything that happened during happened during happened during that session. This that session. This that session. This session is essentially session is essentially session is essentially your unit of your unit of your unit of work. The work. The harness also worked harness also worked harness also worked inside the container inside the container , and the , and the sandbox execution also sandbox execution also sandbox execution also happened happened happened inside the container inside the container . The advantage was that . The advantage was that . The advantage was that since everything since everything since everything was in a was in a was in a container, there was no container, there was no container, there was no need to need to need to design design design service boundaries. Everything was service boundaries. Everything was service boundaries. Everything was local, easy to local, easy to local, easy to edit, and edit, and edit, and editing files editing files editing files felt like felt like felt like system calls. But it so system calls. But it so system calls. But it so happened that we happened that we happened that we realized: we had a realized: we had a realized: we had a " " " pet." If you pet." If you pet." If you haven't heard the analogy of haven't heard the analogy of haven't heard the analogy of "pets "pets " and "livestock" from the " and "livestock" from the " and "livestock" from the old old old infrastructure days, a " infrastructure days, a " pet" is when pet" is when pet" is when you have a named you have a named you have a named entity that you entity that you entity that you manually care for.

manually care for. manually care for. When he is sick, you When he is sick, you When he is sick, you must nurse must nurse must nurse him back him back him back to life. You simply to life. You simply to life. You simply cannot afford to cannot afford to cannot afford to lose it. And lose it. And lose it. And "livestock" is something "livestock" is something "livestock" is something more interchangeable, more interchangeable, more interchangeable, disposable. When one disposable. When one disposable. When one unit unit unit fails, you replace it with fails, you replace it with fails, you replace it with another and keep going another and keep going . A good example: . A good example: . A good example: imagine you have a imagine you have a imagine you have a single database layer. single database layer. single database layer. If it falls, your If it falls, your If it falls, your product doesn't work—it's product doesn't work—it's product doesn't work—it's your "favorite." But your "favorite." But your "favorite." But imagine that you have imagine that you have master- replica replication set up. The master replica replication set up. The master replica replication set up. The master falls and you can falls and you can falls and you can pick up the replica. That's what pick up the replica. That's what pick up the replica. That's what "cattle" is, right? "cattle" is, right? "cattle" is, right? You have many You have many You have many instances, one instances, one instances, one fails—it's fails—it's fails—it's not a problem, you not a problem, you not a problem, you can can can restore it. So, our restore it. So, our restore it. So, our initial setup initial setup initial setup was essentially a " was essentially a " pet" setup. There pet" setup. There pet" setup. There were several problems with this. were several problems with this. were several problems with this. Debugging was Debugging was Debugging was very difficult very difficult very difficult because everything was because everything was because everything was running in a running in a running in a container. We didn't container. We didn't container. We didn't know what exactly was know what exactly was know what exactly was causing the problem. A causing the problem. A causing the problem. A lost packet or a lost packet or a lost packet or a stopped container stopped container stopped container looked almost looked almost looked almost the same, and we the same, and we the same, and we had to had to had to connect via connect via connect via WebSocket to the container WebSocket to the container WebSocket to the container for debugging. for debugging. Another problem Another problem was that was that was that the investigation was the investigation was the investigation was difficult because these difficult because these difficult because these containers containers containers stored stored stored customer data. So we customer data. So we customer data. So we essentially had no essentially had no essentially had no way to get way to get way to get inside to inside to inside to debug. This was debug. This was debug. This was another issue another issue another issue we discovered during we discovered during we discovered during Early Access. And Early Access. And Early Access. And lastly, if a customer lastly, if a customer lastly, if a customer wanted us to wanted us to wanted us to connect to, connect to, connect to, say, an MCP in say, an MCP in say, an MCP in their VPC, that their VPC, that their VPC, that meant peering the meant peering the meant peering the networks or sending networks or sending networks or sending them our harness to them our harness to them our harness to run on their run on their run on their own. So, there own. So, there own. So, there were a lot of problems with were a lot of problems with were a lot of problems with that architecture that that architecture that that architecture that we couldn't we couldn't we couldn't find a really find a really find a really good solution for. So good solution for. So good solution for. So we decided that we we decided that we we decided that we wanted to essentially wanted to essentially wanted to essentially separate the "brain" separate the "brain" separate the "brain" from the "hands," and I'll explain from the "hands," and I'll explain from the "hands," and I'll explain what that means. The agent

what that means. The agent what that means. The agent essentially turned essentially turned essentially turned into three separate into three separate into three separate interfaces. On the left interfaces. On the left interfaces. On the left you have the session—this is essentially a you have the session—this is essentially a log just for log just for log just for adding everything that adding everything that adding everything that has happened in the session up to has happened in the session up to has happened in the session up to that point. that point. that point. In the middle is the "brain", In the middle is the "brain", In the middle is the "brain", which is your framework or which is your framework or which is your framework or agent cycle along with agent cycle along with agent cycle along with Claude. And on the right—" Claude. And on the right—" hands." That is, hands." That is, hands." That is, sandboxes, your sandboxes, your MCP tool challenges, etc. They MCP tool challenges, etc. They are created on are created on are created on request, disappear request, disappear request, disappear when you're done when you're done when you're done , and if you , and if you , and if you lose the sandbox, lose the sandbox, lose the sandbox, you just start you just start you just start it again. So each it again. So each it again. So each component became component became component became interchangeable and " interchangeable and " non-unique". When non-unique". When non-unique". When the sandbox "died" the sandbox "died" , the framework or Claude could , the framework or Claude could , the framework or Claude could just start it just start it just start it again, right? again, right? again, right? The error returned The error returned The error returned to Claude, and for him the to Claude, and for him the to Claude, and for him the sandbox was sandbox was sandbox was essentially the same essentially the same essentially the same tool as tool as tool as any other. If the any other. If the any other. If the tool call tool call tool call fails, the sandbox fails, the sandbox fails, the sandbox "crashes" and it "crashes" and it "crashes" and it decides what to do: decides what to do: decides what to do: run it again or run it again or run it again or move on. When move on. When the framework "crashed", you could the framework "crashed", you could the framework "crashed", you could start a new start a new start a new instance and instance and instance and continue where continue where continue where you you you left off, simply left off, simply left off, simply by reading the by reading the by reading the session event log. Work session event log. Work session event log. Work resumes from the resumes from the resumes from the last event as last event as last event as if nothing had if nothing had if nothing had happened. Nothing happened. Nothing happened. Nothing happened. The only thing we were happened. The only thing we were happened. The only thing we were protecting here was protecting here was protecting here was your session log.

your session log. your session log. Everything else was Everything else was Everything else was interchangeable: it interchangeable: it interchangeable: it could fail and could fail and could fail and be easily be easily be easily repaired. Another thing repaired. Another thing repaired. Another thing we did was to ensure that account we did was to ensure that account we did was to ensure that account data never data never data never entered our entered our entered our sandbox. In our sandbox. In our sandbox. In our previous system, previous system, previous system, your data was your data was your data was sandboxed, so you sandboxed, so you sandboxed, so you were more vulnerable were more vulnerable were more vulnerable to prompt injections. to prompt injections. We changed it this way: We changed it this way: if the sandbox if the sandbox if the sandbox needed to needed to needed to access something access something access something that required a token or that required a token or that required a token or access control, access control, access control, the request went through a the request went through a the request went through a proxy that appended proxy that appended proxy that appended your token before your token before your token before calling the external calling the external calling the external endpoint or MCP. endpoint or MCP. This way, the model This way, the model never sees never sees never sees your your your credentials or tokens, credentials or tokens, credentials or tokens, so you can't so you can't so you can't force Claude through force Claude through force Claude through prompt injection prompt injection prompt injection to say, for example, " to say, for example, " Give me your tokens Give me your tokens ." Either display the tokens ." Either display the tokens ." Either display the tokens on the screen or write on the screen or write on the screen or write them to a file. You will them to a file. You will them to a file. You will never be able to never be able to never be able to do this because Claude do this because Claude do this because Claude never actually never actually never actually sees your tokens. sees your tokens. Because we had an append-only event log, our data compaction strategy remained data compaction strategy remained flexible. So, nothing is flexible. So, nothing is flexible. So, nothing is ever ever ever deleted. We do not deleted. We do not deleted. We do not compress or compress or compress or change your flow of change your flow of change your flow of events. We always events. We always events. We always store every store every store every event that happens, and the event that happens, and the event that happens, and the compression strategy compression strategy compression strategy decides, “Oh, do I want to decides, “Oh, do I want to decides, “Oh, do I want to go through the entire go through the entire go through the entire event log, event log, event log, summarize it, and summarize it, and summarize it, and add context?

add context? add context? Do I want to review Do I want to review Do I want to review only the last n events and only the last n events and only the last n events and add them to the add them to the add them to the context? So, context? So, context? So, thanks to the interface thanks to the interface thanks to the interface design we chose, design we chose, design we chose, we could we could we could have a more flexible have a more flexible have a more flexible context strategy. context strategy. And another benefit And another benefit we noticed: we noticed: we noticed: because we could because we could because we could start the start the start the logical logical logical inference calls immediately, we were able to inference calls immediately, we were able to inference calls immediately, we were able to reduce the time to the reduce the time to the reduce the time to the first token by first token by first token by about 90%. So about 90%. So about 90%. So Claude Claude Claude only started the sandbox when only started the sandbox when only started the sandbox when needed, needed, and didn't wait for the and didn't wait for the container or container or sandbox to be deployed for you to sandbox to be deployed for you to sandbox to be deployed for you to get your first get your first get your first token. So thanks to token. So thanks to token. So thanks to this architectural this architectural this architectural change, we were able to change, we were able to change, we were able to get a much get a much get a much faster and more responsive faster and more responsive faster and more responsive system. And here is system. And here is system. And here is an illustration of how an illustration of how an illustration of how one "brain" can one "brain" can one "brain" can control many " control many " hands." Your hands." Your hands." Your loopback agent system loopback agent system loopback agent system can now can now can now access a access a access a container in container in container in the sandbox, a the sandbox, a the sandbox, a customer's VPC, or an customer's VPC, or an customer's VPC, or an MCP server. So all MCP server. So all MCP server. So all these "arms" are essentially these "arms" are essentially these "arms" are essentially interchangeable, you interchangeable, you interchangeable, you can connect can connect can connect many of them, which many of them, which many of them, which made our made our made our architecture much more architecture much more architecture much more flexible. If you flexible. If you flexible. If you remember only remember only remember only one thing, let it be that one thing, let it be that one thing, let it be that systems systems systems contain assumptions, and contain assumptions, and contain assumptions, and assumptions can become assumptions can become assumptions can become outdated. So we outdated. So we outdated. So we designed interfaces designed interfaces that would survive these that would survive these that would survive these assumptions and made assumptions and made assumptions and made the architecture more the architecture more the architecture more flexible. This allowed flexible. This allowed flexible. This allowed us, as I said, us, as I said, us, as I said, to change the to change the to change the compression strategy, compression strategy, compression strategy, to completely replace the system, but the to completely replace the system, but the to completely replace the system, but the interface and interface and interface and architecture architecture architecture remain remain remain the same, and we the same, and we the same, and we can work with can work with can work with it in a more modular way.

it in a more modular way. Okay, now we'll Okay, now we'll move on to the move on to the move on to the practical part. practical part. We've prepared We've prepared free API free API credits for you to credits for you to credits for you to help you help you help you build with us build with us build with us today, so today, so today, so please scan the QR please scan the QR code, log in to your code, log in to your code, log in to your Cloud Cloud Cloud Platform account or create a Platform account or create a Platform account or create a new one, and you can new one, and you can new one, and you can get those credits get those credits get those credits to build with to build with to build with me and then with me and then with me and then with Gabe. And I'll leave that Gabe. And I'll leave that Gabe. And I'll leave that for a moment. Greetings. for a moment. Greetings. Excuse me, what? Last Excuse me, what? Last slide. That's actually what slide. That's actually what slide. That's actually what we were talking about here we were talking about here we were talking about here , right? These are kind of the , right? These are kind of the , right? These are kind of the three three three interfaces that interfaces that interfaces that we designed the we designed the we designed the system for, right? The first system for, right? The first system for, right? The first question was question was question was what interfaces did what interfaces did what interfaces did we develop to we develop to we develop to keep the system keep the system keep the system relevant despite relevant despite relevant despite the changes? So we have a the changes? So we have a the changes? So we have a session log just for session log just for session log just for appending, there's a framework appending, there's a framework appending, there's a framework that that that can be replaced again, and can be replaced again, and can be replaced again, and tools where tools where tools where you can plug in a you can plug in a sandbox, sandbox, sandbox, tool calls, or tool calls, or tool calls, or anything that anything that anything that the model might the model might the model might need access to— need access to— this makes the system this makes the system this makes the system more versatile and more versatile and more versatile and extensible. Okay, extensible. Okay, extensible. Okay, back to back to the credits. Does the credits. Does anyone else need to anyone else need to anyone else need to scan this before scan this before scan this before we start? Okay, we start? Okay, we start? Okay, I still see a few I still see a few I still see a few phones. Okay, I'll give phones. Okay, I'll give phones. Okay, I'll give everyone about 30 everyone about 30

everyone about 30 seconds. Thank you for your seconds. Thank you for your seconds. Thank you for your patience. Oh, are there any patience. Oh, are there any patience. Oh, are there any problems with the problems with the problems with the internet? Oh no. internet? Oh no. Gabe, do we have Gabe, do we have a microphone? Because I can't hear a microphone? Because I can't hear a microphone? Because I can't hear some questions. Yes, some questions. Yes, some questions. Yes, the credits should the credits should the credits should arrive as soon as you arrive as soon as you arrive as soon as you sign up, sign up, sign up, hopefully within a hopefully within a hopefully within a minute or two. We are minute or two. We are minute or two. We are now taking now taking now taking the microphone so that we can the microphone so that we can the microphone so that we can walk around the hall and walk around the hall and walk around the hall and answer these answer these answer these questions. But I'll questions. But I'll questions. But I'll show you how to show you how to show you how to create a very create a very create a very simple agent that simple agent that simple agent that tells jokes. I'll tells jokes. I'll tells jokes. I'll first talk first talk first talk about our core APIs and about our core APIs and about our core APIs and concepts, and then we'll concepts, and then we'll concepts, and then we'll go to go to go to platform.anthropic.com to platform.anthropic.com to platform.anthropic.com to create your first create your first create your first agent. Our first agent. Our first agent. Our first step is to create an API for step is to create an API for step is to create an API for environments (V1 environments), environments (V1 environments), environments (V1 environments), which helps which helps which helps create an environment create an environment create an environment or template for your or template for your or template for your sandbox. Here sandbox. Here sandbox. Here you will find you will find you will find your network settings. Do you your network settings. Do you your network settings. Do you want unlimited want unlimited want unlimited network access or do network access or do network access or do you want to limit it to you want to limit it to you want to limit it to specific domains or specific domains or specific domains or MCPs? You can MCPs? You can MCPs? You can specify which specify which specify which packages should be packages should be packages should be installed in your installed in your installed in your container or " container or " sandbox". So, this sandbox". So, this sandbox". So, this will be our first step will be our first step will be our first step in setting up the in setting up the in setting up the environment. Next, you environment. Next, you environment. Next, you create an agent.

create an agent. create an agent. An agent is essentially an An agent is essentially an An agent is essentially an object that can be object that can be object that can be given a name. In this given a name. In this given a name. In this case, we will call case, we will call case, we will call him a "joker." him a "joker." him a "joker." We will choose the model We will choose the model We will choose the model you want to you want to you want to use for use for use for this agent. You this agent. You this agent. You can can can use Opus, use Opus, use Opus, Haiku, or Sonnet—your Haiku, or Sonnet—your Haiku, or Sonnet—your choice. After that, we choice. After that, we choice. After that, we will provide the agent with a will provide the agent with a will provide the agent with a system prompt. In system prompt. In system prompt. In this case, we would this case, we would this case, we would say: "You're a comedian, say: "You're a comedian, say: "You're a comedian, find a fresh joke on a find a fresh joke on a find a fresh joke on a given topic and given topic and given topic and tell it well." Then tell it well." Then tell it well." Then you can give you can give you can give that agent that agent that agent tools. We will tools. We will tools. We will provide him with the provide him with the provide him with the agent toolkit, agent toolkit, agent toolkit, which is our basic which is our basic which is our basic set. It's like your set. It's like your bash toolkit: read, write, grep, bash toolkit: read, write, grep, etc. Almost all etc. Almost all etc. Almost all agents will agents will agents will need this. This will be need this. This will be need this. This will be your second step. The your second step. The your second step. The third step third step third step will finally be will finally be will finally be to create a session. So to create a session. So , the session takes the , the session takes the , the session takes the agent ID agent ID agent ID and the and the and the environment ID and creates environment ID and creates environment ID and creates the session. This is your the session. This is your the session. This is your task, i.e., a task, i.e., a task, i.e., a unit of work or unit of work or unit of work or your assignment. And when your assignment. And when your assignment. And when the session is initialized the session is initialized , it is in , it is in , it is in standby mode. standby mode. standby mode. No work has No work has No work has been done yet, right? been done yet, right? been done yet, right? Until you send Until you send Until you send the event. And this is what the the event. And this is what the flow of flow of events in your session will look like events in your session will look like . You will have events . You will have events . You will have events that you send, that you send, that you send, as well as events that the as well as events that the as well as events that the agent will agent will agent will respond to you with.

respond to you with. respond to you with. On the left you will see On the left you will see On the left you will see user.message. You can user.message. You can user.message. You can say, "Tell say, "Tell say, "Tell me a joke about me a joke about me a joke about distributed systems distributed systems ." And then the agent ." And then the agent ." And then the agent will answer you with a will answer you with a will answer you with a funny joke. You funny joke. You funny joke. You can can can use what is use what is use what is called a called a called a custom custom custom tool. A tool. A tool. A custom custom custom tool essentially tool essentially tool essentially allows you to define the allows you to define the allows you to define the tool you tool you tool you want to want to want to use to use to use to handle how you handle how you handle how you respond to it. And Claude will respond to it. And Claude will respond to it. And Claude will essentially add a essentially add a custom custom tool usage event to tool usage event to tool usage event to your event stream, your event stream, your event stream, which you can which you can which you can track, track, track, process, and process, and process, and return the result to return the result to return the result to Claude as a Claude as a Claude as a custom custom custom tool result. And there are a tool result. And there are a tool result. And there are a few other few other few other examples of the examples of the examples of the types of events you might types of events you might types of events you might see. So, you see. So, you see. So, you will see session. error will see session. error will see session. error if your session if your session ever throws ever throws ever throws or or or encounters an error. encounters an error. encounters an error. Idle status is when Idle status is when Idle status is when your session is idle and your session is idle and your session is idle and not working. not working. not working. This usually happens when This usually happens when This usually happens when you're not actually you're not actually you're not actually using using using the sandbox. And interrupt on the sandbox. And interrupt on the sandbox. And interrupt on the left is when you the left is when you the left is when you want to stop or want to stop or want to stop or interrupt the session and interrupt the session and interrupt the session and no longer want to no longer want to no longer want to continue working. continue working. This is kind of our This is kind of our complete list of complete list of complete list of features. Maybe features. Maybe features. Maybe if there's time at the end if there's time at the end if there's time at the end , Gabe , Gabe , Gabe can come back to can come back to can come back to it and dig deeper into the it and dig deeper into the it and dig deeper into the topic, but Gabe topic, but Gabe topic, but Gabe will work with you will work with you will work with you on his part of the on his part of the on his part of the workshop, looking more closely at workshop, looking more closely at concepts like concepts like multi-agency and multi-agency and multi-agency and memory, which we also memory, which we also memory, which we also provide. A. Okay.

provide. A. Okay. provide. A. Okay. Before we get to Before we get to Before we get to that, we'll now that, we'll now that, we'll now go to go to go to platform.claude.com. platform.claude.com. platform.claude.com. I hope you I hope you I hope you can use can use can use your API credits and your API credits and your API credits and see them now. see them now. see them now. If not, Travis If not, Travis If not, Travis walks around with a walks around with a walks around with a microphone, so you microphone, so you microphone, so you can reach out to can reach out to can reach out to him if you have any him if you have any him if you have any questions questions questions or need or need or need help. And if you help. And if you help. And if you go to your go to your go to your console, in the console, in the console, in the agent management section agent management section agent management section you will see something you will see something you will see something called "quick called "quick called "quick start". This is kind of start". This is kind of start". This is kind of our guide on our guide on our guide on how to create your how to create your how to create your first agent.