Building AI agents with Claude in Amazon Bedrock

Introduction and Overview 00:06

  • Presenters from AWS introduce the topic: building AI agents with Claude models via Amazon Bedrock.
  • Amazon Bedrock is described as a fully managed service that gives access to foundational AI models through a unified API, providing scalability, model choice, guardrails, and enterprise-grade security.
  • "Agents" are defined as autonomous systems that can reason, plan, and perform multi-step tasks to achieve objectives, similar to human workflows.

Introduction to Strands Agent SDK 02:53

  • AWS has released an open-source SDK called "Strands Agent" for building agentic AI applications.
  • Strands Agent SDK focuses on simplicity: it requires only models, tools, and prompts to create an agent.
  • The SDK lets the large language models (LLMs) handle more reasoning and actions, providing model flexibility and reducing the need for extensive code scaffolding.
  • Strands comes with built-in tools and supports easy deployment across AWS services like EC2, Lambda, and ECS.
  • The SDK, documentation, and GitHub links are shared, and community contributions are encouraged.

Workshop Setup and Getting Started 05:44

  • Participants are provided preconfigured AWS accounts and browser-based VS Code environments for hands-on practice.
  • Step-by-step guidance is given on signing in, accessing the AWS console, and enabling relevant Claude models in Bedrock.

Demonstrating Strands Agent Capabilities 06:35

  • A video example demonstrates using Strands Agent to generate a math animation video (similar in style to "3Blue1Brown") by scripting a quadratic equation plot, powered by Claude 3.7 and Strands.
  • All code and workshop materials are open source and available for replication.

First Exercise: Weather and Word Count Agent 10:44

  • The initial workshop exercise involves building an agent that retrieves weather data and counts words in the response using Strands.
  • Demonstrates importing tools, creating system prompts, and leveraging the HTTP request tool (which is built-in and requires no API keys).
  • Models can be swapped easily (e.g., Claude 3.5), and custom tools can be created with simple decorators.
  • Querying "What's the weather like in San Francisco?" shows the agent fetching data and performing word count in about 44 lines of code.

Q&A: Strands Agent Features 13:49

  • Strands agent is highlighted as easy to use compared to other agentic frameworks due to minimal boilerplate and simple tool creation.
  • The model autonomously determines how to extract geographical data (e.g., latitude, longitude) from queries and interact with APIs.

Second Exercise: Using MCP Servers for Documentation and Diagrams 15:29

  • Introduces MCP (Model Connection Protocol) servers, which facilitate complex tasks like searching documentation or generating AWS diagrams from prompts.
  • Example demonstrated: an agent queries AWS Lambda documentation and creates a static website architecture diagram, interacting with multiple MCP servers in about 40 lines of code.
  • The workflow breaks tasks into steps, using HTTP calls to search, read, and visualize, with results varying based on context.

Q&A: Hosting MCP Servers on AWS 19:47

  • MCP servers can be hosted behind API Gateway and Lambda functions for server-side event streaming; open-source code samples are available for cloud deployment.

Third Exercise: Building a CDK Agent with Claude Code 20:30

  • Demonstration of creating an agent for AWS CDK (Cloud Development Kit) tasks via Claude-powered code assistance.
  • Claude Code assists in creating a new Strands agent that connects to an MCP server by referencing prior files and generating the needed code.
  • The agent is run, and it generates step-by-step CDK code examples (e.g., to create an S3 bucket), showing the agent's ability to guide and produce code automatically.
  • Claude Code reports on API cost, runtime, and changes made during the process.

Use Cases and Further Q&A on MCP Servers 24:33

  • MCP servers provide external context, allowing agents to look up documentation, perform cost analysis, draw AWS diagrams, connect to databases, and more.
  • They act as "USB-C" connectors for LLMs, enabling access to various resources such as documentation, cost tools, architecture diagram tools, and infrastructure as code integrations (CloudFormation, Terraform).
  • Multiple AWS-official MCP servers exist, catering to functions like cost analysis, front-end code, database interaction, and drawing.

Closing and Credits 26:54

  • Participants are offered $25 in AWS credits for feedback.
  • The presenters encourage experimentation with Strands, MCP servers, and Claude in Bedrock, emphasizing open source and community involvement.