Building AI agents with Claude in Amazon Bedrock | Code w/ Claude

Building AI agents with Claude in Amazon  Bedrock. Today I am excited to explore how   to create intelligent autonomous AI systems that  can transform your applications. My name is Dwan   Lightoot. I am a developer advocate at AWS.   I'm Banjo. I'm a systems architect at AWS. Hi   everyone. My name is Suman DNA. I'm a developer  advocate at AWS. Now, this will be a hands-on   keyboard event. So, we have a live workshop where  you can log into an environment and get started   with some code. So, we're going to go through a  few slides to kind of level set and get everyone   on the same page and then we'll hop into the  workshop. Now this is code with Clyde and one of   the things that I want to talk about is enthropic  on AWS. To do that we have Amazon bedrock which   which is a fully managed service that provides you  access to powerful foundational models like those   in the cloud family through a unified API. This  gives you everything you need to not only build   but also scale your AI applications globally. And  it does this by providing you with everything you   need model choice guard rails as well as it gives  you security at enterprise grade at default. Now   we're talking agents. That's what everyone is  here to see. But we kind of need to level set   what that really means. And at AWS, we explain it  like this. An agent is an auton autonomous system   that can reason, plan, and take multiple  steps to perform an objective like humans.    So if you have a task and you give it to an agent,  the agent is able to take that task and say here   is the highle objective. Let me create a plan of  the steps that I need to take. Then it could take   actions on those steps. And once those steps have  been taken on those steps, then it can evaluate   the results and reason on what needs to happen  next until it actually achieves that objective.    This is an agitic system that we'll be discussing  today. Now at AWS, we've also done something   awesome that my colleague Suman is going to talk  about at this time. Thank you so much Dan. So   taking one step further what we have done in  fact last week last Friday we have announced   an open-source SDK to build agentic application  called strands agent. So what this agent does is   it's a very simple SDK which needs three things  models tools and prompt. It cannot go any simple   than that. It doesn't have any scaffolding  that you don't have to guard rail your prompts   um your backstory goals etc. What we believe in is  that in today's world the LLMs are pretty strong   and we want to make use of the full strength  of the model in the back end. So we are giving   the better flexibility for the model to reason on  behalf of us. And that's why if you look at this   architecture it's very straightforward. You create  a prompt or this is your question. You send it to   the agent which is the strands agent. And when you  create an agent object, you will define the model   and tools that we provide to you. You will see in  the workshop when Banjo will go through a couple   of the demos, you will see that the default  model that we use is claw 3. 7 as of today.    And the moment you have strands installed and  configured, you get a bunch of tools inbuilt.    So you don't have to write heavy lifting code to  create few of your basic needs. And the best part   is not only for your testing and deployment in  the test environment, you can actually deploy on   the cloud. So assume that you have your workforce  in EC2 or Lambda or ECS, you can just deploy your   code with ease uh with a integrated uh support  for all these services. So to get started just

uh take a snap of this. This contains the launch  blog as well as the documentation which is nothing   but strands. com and the GitHub. And this is  an open-source uh project. So feel free to   uh give a star and uh you know just uh raise  a PR if you have anything interesting that you   build or if you have any specific requirement  which is not there uh feel free to contribute.    It's just 3 days old and we got many PR and uh  feedback from the community and we would love   to work with all of you. All right so the moment  you're all waiting for. it's time to build. So,   we have pre preconfigured AWS accounts for you.   Uh, VS Code is also configured. So, there's   nothing you need to download. Everything is done  in the browser. So, when you go to this URL, it's   going to take you to a popup screen and you're  going to uh sign in. You're going to it's going   to ask you for a one-time password. Uh, that's the  best way to log in. And once you do that, you're   going to be able to have access to the AWS account  as well as the VS Code server we spun up. So,   uh, we're going to take a few minutes to get set  up. I always say this is the hardest part of the   workshop. So, uh, Tuman and Dan are going  to be walking through. Ask any questions,   raise your hand, and I'll also walk through it.   But, uh, take some minutes to get started and,   uh, let's get start building. So, when you log in,  you'll get to a screen like this. So, I'll give   a couple more minutes to get to the screen.   Then, I'll walk through setting up Bedrock so   we can have access to the models and then to our  VS Code server. Well, while people are waiting,   I have a video to highlight what strands can do.   So, I can just show that in the meantime. Uh,   so this example, we actually have a strands agent  that's actually going to uh create a math video   for this. So, it's actually pretty cool. So, let's  see what happens here. So, it's running an MCP   server. It's going to start that uh requirements.   It's going to actually create an animation video   for us. So it's going to create a maximum scene  that draws a cubic function. Uh 2x cubed uh yeah   something hard that you might have to  do in latex scientific like this is very annoying   but we'll see how the MCP server is actually going  to build it and actually make a video to highlight   it. How many of you have uh heard of three blue  one brown? So that is what you are going to see   now. So we have created an MCP server which can  create the videos that you see in three blue one   brown. So we have created a quadratic equation and  we wanted to plot that within the range of minus3   to 3 and this is powered by cloud 3. 7 and strands.   So uh we'll push that code in the GitHub repo   which we have shared earlier but uh this is just  a testimony of how you can get started quickly   with the out of the box tools and just few lines  of code. Thank you. So the first thing to do uh   is make sure you log into this AWS account. So you  should not be using your own AWS account. We have   already provisioned one with all the resources  needed. Let's open AWS console and it'll open up   a new account new like this. You should see this  work stop participant role. So you should not be   using your own AWS account if you have it. From  here we're going to type bedrock Amazon Bedrock. So we get to the Amazon Bedrock console  screen. We're going to want to enable   the models so we can use them in our  app in our lab. So give it a second.    I'm going to scroll down to this model access  button and then I click the modify model access   and we got some new models but they have not  enabled them yet. So we're just going to use   claude v7 uh 3. 5 hiq and 3. 5 sonnet. So enable  those. Got the four already. We got four already.    Yes. But I don't think it's enabled in this  account yet. So we got to use the older models.    So and press next. Um once we request access  submit but again this code all the code is   open source the workshop is open source and we can  share the link so you can run this on yourself. So   uh module one is the one we're doing today.   So we're going to show you how strand agent   works and how you can actually build a aentic  workflow with it. So the first thing you do   is install it. So just doing pip install strand  agents and strand agent tools gets you what you   need and have some utility things UV to download  MCP servers. So I'm going to copy that command paste right already installed great so and then the  cool thing with cloud code you actually can use   it of Amazon Bedrock. So if you have an ads  account, you have bedrock, you can use cloud   code without using the entropic key signing in  everything is just done through bedrock. That's

by using this export cloud code uh environment  variable. So let me go ahead and do that. All right. So claude get started research preview.   Gonna push dark mode of course. All right. Cool.    So use recommended settings. Yes, proceed. All  right, cloud code is ready to go. So all I did   was just export that command because it's already  using Bedrock. Everything's just ready to go. It's   already in my environment. And you all can see  this, right? This is a good screen. Yeah. Cool. All right. So the first uh exercise of the  workshop, we're actually going to use a   weather count. Every AI thing has to start with  weather first. So, you know, well, we added two   tools to this one. One that can actually get the  weather and then actually count how many words   are in the response just to show how easy it is to  use different tools using strands. So, and we're   actually going to use cloud code to explain how  this code works as well before I dive into it. So,   I'm going to open this up. Paste. Can you explain  the structure of the strand agent warther world   count file? All right. So then it's going to be  able to see what happens. Use the tokens. Let   me open the file while it does that and we'll  see what Claude says. Let me make this bigger. All right. So demonstrate a simple agent  implementation. the Stram framework agent   file import necessary word count tools using  cloud 3. 5 executed query so let's walk through   the code in more detail so we actually have a  system prompt and we're saying you know find the   weather and it actually puts the API in the system  prompt httpe. gov gov. So there's no API key needed   for this. So I can just query this and get the  actual weather of what the place is and provide   in a human readable way. And the cool thing about  this uh strands has this HTTP request tool already   built into the framework. So it's going to be able  to just make that request for you automatically.    You just provide URL is able to call that and  get the actual data from that. And then you can   you the cool thing about strands also is you  can change the different models you can use.    You can use light lm, you can use lama, uh you  can use bedrock. Bedrock's the default. So I just   changed the claw 3. 5 just to be faster here. Uh  and then the cool thing about strange is that's   the system prompt. The big system prompt we put  up there which tools we're going to use the word   count tool. Now what I really like about the tool  decorator is I just define a function and just put   the return value. There's no crazy things, no  adding. I just put this tool decorator and it   handles the rest. So as a developer is building  functions, I want to put all this extra stuff in   there, make it as simple as possible to have  a tool. So that's a really big plus to the   strands framework. And then from there, I can  just say, what's the weather like in Seattle?    Let's change it. San Francisco and let's go  to the terminal. Oops. Strand agent three. All right. So we can see how it's going here. Uh  first I'll get the coordinates for San Francisco.    You can see the HTTP request tool. Now it's  going to use the HTTP request to find that   weather. It gets the San Francisco 65 sunny west  winds few days highlight. Now uses the word count   tool 110 words. So with about 44 lines of code  we're able to make that API request. We have the   agent. It's going to have multiple tools. Done.   So I'll pause here for a second just to have any   questions of strands because not everybody's  going through it. So Strand is a open source   SDK for building agents. So there are a lot of  agentic frameworks but strand is it own one and   you can see that you really just need a system  prompt the tools and a model and it can execute   that loop. So yeah it is similar to other agentic  frameworks but I believe it's much easier to get   started without the boiler plate and extra  stuff that you've seen in other frameworks.    All right, one more question and we'll move  on to the next part. Uh, thank you. So,   I have a code related question. So, how are you  passing like the latitude, longitude, and zip   code like inside the string? So, so the cool thing  is that we're letting the model decide how to do   that. It uses this HTTP request thing and it  understands what the latitude and longitude is of   San Francisco and is able to pass that to this API  endpoint or use API zip code. So it understands   the what the API is based on the system prompt  and the model is figuring that out by itself.    So we're handing a lot of the infrastructure you  might see in other agent frameworks saying you got   to do this, you got to do this. We're letting  the model decide to do that because the models   are much more capable than they were two years ago  when you saw the first type of agentic frameworks

coming out. So this trans tools uh like I see  that you're importing two tools uh which is like   word count and HTTP request. So how many tools are  there like? Yeah. So there's some built-in   tools in the strand framework like HTTP request  but then I also just made my own tool which is   literally this one line of code return this lens.   So it's very easy to make your own custom tools.    You just put this tool decorator and that's it. So  very streamlined. All right. So we're going   to move on to the next exercise. Uh so the next  exercise is fun one. MCP servers. So you know MCP   is the hottest thing you know. So, uh, what's  really cool about strand, it has built-in MCC   support and AWS, we actually have official AWS,  uh, MCP servers. So, I'm going to highlight the   documentation lookup because AWS can be very long,  a lot of different things. So, if I just have one   endpoint to just grab that information and pass it  to the model, it's going to be able to understand   how to build. And there's an architecture diagram.   So, making AWS diagrams. So there are two MC   there's a whole bunch of them listed here but I'm  going to highlight these two an example. So let's   go back and we can also ask claude code to explain  it as well. So get back me open my cloud code. So I have to explain how the MCP  diag one works. So let me open   that up while it's waiting. Let me close that. Cloud code's documenting. It's looking what to do. It's taking it taking its time. So while cloud code is going, I'll explain  what's happening. That just finished. Cool.    All right. Import the strand. Still going. All  right. So the cool thing about uh MCP server is   we can actually just pull it in with one line.   We just the command uvx. This is uh the Python   material for downloading things. So we can just  point it to the MCP server and it'll download it   locally. So I'm passing in uh the documentation  MCP server and I'm passing in the diagram MCP   server and then you know claude even saying  the same thing connect to the MC diagram run   one configure using Bedrock again I can choose  different models on Bedrock to pass in and then   I'm giving a system prompt you an extra certified  solutions architect your role to help customers   understand best practices query documentation  generate diagrams tell the customer the full   path of the diagram when you create it just so  we know where it is and then again we just Okay,   we pass in multiple MCP servers with the diagram  client with the docs client. Uh we get all the   tools based in those uh MCP servers and then  again you know we have the tools we have the   model and the system prompt. All you need to have  the stand agent and then we give it like get   the documentation for AWS Lambda then create a  diagram of a website that uses lambda for a static   website hosted on S3. So that's the task I give  it. Let's see how it executes that. So oops. All right. So downloads the MCP server. It's already  downloaded that already and now it's processing.    Now it breaks it down into steps. So first  I'll search the AW documentation. Then I'll   read the documentation. Then I'll create a diagram  illustrating architecture. So you can actually see   it stop process. It makes an HTTP call to get the  the search. It finds the Lambda welcome page to   get the information. It queried the documentation.   Now it's going to draw the diagram. Let's see. It listed the icon. So what icons are available  for it. Now it's actually generating the diagram.    So excellent. It failed there, but it understood  that and explained. Let me correct the cloud icon.    Now it generated it, explained what's happening,  and it made this new diagram saved here. So I   should be able to open it up. Generated diagrams.   And tada, it made a new diagram for me. So,   whoops. Let me Yeah, it's going to be different  every time when you of course based on how much   context you give it. But yeah, I'm able to  make that diagram and just about, you know,   40 lines of code. I have two MCP servers. I put  my system prompt picking a different model and I'm   able to create that agentic workflow very quickly  and very easily without any too much headache.    So I'm going to pause here and he talks about MCP  servers connecting maybe with a strange agents.    Is there a pattern um through API gateway  to host a MCP server with lambda like the

server side events? Is there something that  allows Yeah. Now because MCP supports   the HTTP streaming server side event you can  have it in a Lambda function and do that. There   are some I believe there are some open source  code that demonstrates how to do that. One of   my colleagues has done that. So yes, totally  possible to do that use case because right   now I'm just running the MCP ser locally,  but if you wanted to have it in the cloud,   I got a lambda function. Totally possible use  case. You would have to change how this is set up. All right. So the last exercise was  we're actually going to create a new   uh a new agent using cloud code to understand  how it does uh without just looking at the code   we have already and able to understand which CDK  I'm going to make a CDK agent. So CDK is a cloud   development kit. So it's a way to create uh AWS  infrastructure through code. Uh normally if you   want to create as infrastructure you might use  something like cloud formation which is a YAML   based way but if it's a developer uh CDK is more  preferred for that because you can integrate the   python typescript code. So I'm going to  show an example of how we can use claude   code to actually create a new strand agent  for us. So going to make this new file here. Oops. All right. So I created a brand new file, nothing  in it. And then I'm going to ask claude code, you   know, update the CDK agent to create a stand agent  connecting to an MCP server. Uh look at the other   files to understand how to do this. So I'm not  giving you the documentation. I'm just going to   say, you know, use the knowledge you have already  to understand how to make it. So, let me get into the repo and then claude and ask it. So, yeah. So now cloud code  is going to look through. We'll see   the plan of attack it does. So let's  let's see what it's doing. All right. I'll update the CDK agent to CL agent  connected to MCP server. You know it   looks at the other one. There's no one  line of code there. It read the other   word count one. All right. It finds the MCP  doc one. Examines it. Cross it off the list.    All right. On to the next one. Create a stand  agent. It put all the code for that. It asks   me do I want to make this edit. So say yes,  make the edit. It's adding the code to me. All right, configure ready. Then MCP client  configure system prompt. It's done it. Perfect.    So close out of it. And what I like about cloud  code actually tells us how much it costs when   you end it. So the total cost, how long the  API took, you know, how many code changes,   how which models it used, it use cloud 3. 5  site haiku, use cli. So uh using cloud code   of bedrock is that's an easy way to get started  in building applications like this and it's now   natively inside VS code which is very exciting.   So let's see if the code actually works. So   uh I'm just going to run it. Let's see you're  expert AWS CDK expert. Uh how can I get a simple   S3 bucket with CDK? Give us something simple.   Let's see how it works. Python 3 CDK agent. All right. So, I was making a TypeScript example.    I'll provide step-by-step guidance. It  wrote the code for me. It's doing that security text this add a CNN nag rule. So, it  really understands what CDK is. It's putting   all this extra stuff there. Uh, and yeah, it  even explained that. So cloud code is able   to understand how strand agent works create  a new agent based on that create a template   on that and it was very easy to understand. So by  providing the enough context cloud code was able   to understand what to do create this agent and  it's pretty good job. So very impressed with using   cloud code and bedrock power that's also amazing.   Uh but that was kind of the workshop through and   through kind of to show you what strands can do  from a simple weather agent to connecting to MCP   servers and then also making your own agents. So,  I'm going to pause here for any questions or if   you want to see a demonstration of something cool,  happy to try it out and see what Claude can do.    Uh, sorry, I just have to ask like what are some  of the use cases for like using MCP? Like I know   I will be building agents that probably connect  to the outer world like through APIs and all like   uh maybe I don't know a lot about MCPs. So, but  like what are some of the use cases you can like   use MCPS for? Right. So, a great question. What  are the use cases for MCP? So by default the agent   doesn't have enough information to ex connect  to external API. So you saw in the example we   created an AWS diagram. If I did that without an  MCP server the model will not know how to make

the AWS diagram. It won't know how to look  up the AWS augumentation. So it's providing   extra context to the model to do extra action.   So when it comes to using large things model   context is really what empowers what's going  to go on. MCP provides a structured way to   grab that information so it can take action,  read documentation, uh connect to external   applications. So MCP is really kind of the USBC  they call it of connecting the LLM. So for me   it's a great way to just connect uh external  applications get different context for your model and I'll show you some highlights of the AWS MCP   servers just so you can get  a example of what we have. So there are official list of AWS MCP servers that  do a bunch of different things. So I'll go through   some examples. uh show the documentation one uh  bedrock knowledge base if you have a bunch of PDFs   or documents you want to give to an LLM uh cost  analysis so you want to do AWS cost analysis you   want to get a another example uh the MCP server  could be useful for you uh Amazon Nova Canvas if   you want to draw images uh Nova canvas MVC server  the diagram one that we use today you can also   draw AWS architecture diagrams cloud formation so  there's a bunch of different ones and Terraform   if you use Terraform form uh front-end code you  want to specialize in react or using ads amplify   etc. So there's a whole source of M MCP  servers. So one big use case is I want my model   to understand these things. You have a bunch of  documentation Postgress database, Amazon Neptune   server like so this is really the power of MCP. So  there's so many integrations you can have to give   that model extra abilities to do. Uh we do have  AWS credits for you. So if you want to give us   some feedback and surveys and once you fill this  out, you'll get uh we have $25 in AWS credits to   use. So you can play around, try some things out.   Uh, but thank you for coming and let's go build.