LangGraph Comprehensive Agent

A LangGraph agent demonstrating multiple AgentCore capabilities in a single deployment: custom Lambda tools, a direct MCP server connection, web browsing, code execution, and conversation memory.

What This Example Shows

• Gateway Tools: Lambda function (calculator) via AgentCore Gateway
• Direct MCP Server: AWS Knowledge MCP connected directly from agent code (public endpoint, no gateway needed)
• Default Browser: Web navigation and content extraction via PlaywrightBrowser
• Default Code Interpreter: Sandboxed code execution via CodeInterpreter
• Memory: Conversation persistence with list_events tool and automatic saving
• Auto-build: No Dockerfile needed -- the framework builds the image from source via buildpacks
• agents/ directory: Agent code lives in agents/index.js, not the project root

Architecture

User Request --> AgentCore Runtime --> agents/index.js
                                        |
                                    LangGraph ReAct Agent
                                        |
                                    Claude Sonnet
                                        |
        +----------+----------+----------+----------+------+
        |          |          |          |          |      |
   Gateway MCP  Direct MCP  Browser  Code Interp. Memory Direct
        |          |                                      Response
   Calculator  AWS Knowledge
   (Lambda)    (public MCP)

Prerequisites

• AWS account with Bedrock model access (Claude Sonnet)
• Enable global.anthropic.claude-sonnet-5 (the default; override via the MODEL_ID env var) in the Bedrock console
• Docker installed
• Serverless Framework v4+
• AWS credentials configured

Quick Start

1. Deploy

serverless deploy

The framework will:

• Auto-build a Docker image from source (no Dockerfile needed)
• Push to Amazon ECR
• Deploy the Calculator Lambda function
• Create an AgentCore Gateway with the calculator tool
• Create AgentCore Memory (30-day expiration)
• Deploy AgentCore Runtime

2. Test

RUNTIME_ARN=<your-runtime-arn> node test-invoke.js

Get the runtime ARN from serverless info.

3. Local Development

serverless dev

How It Works

Configuration

The serverless.yml defines a calculator Lambda tool in the gateway:

ai:
  tools:
    calculator:
      function: calculatorFunction # Lambda tool via gateway
      toolSchema: [...]

  agents:
    assistant:
      memory:
        expiration: 30

The calculator tool is automatically placed in an auto-created default gateway. The AWS Knowledge MCP server is connected directly from the agent code (no gateway needed).

Tool Sources

The agent in agents/index.js combines six tool sources:

1. Gateway tools -- calculator Lambda discovered via BEDROCK_AGENTCORE_GATEWAY_URL using MCP protocol with SigV4 auth
2. Direct MCP tools -- AWS Knowledge MCP server (https://knowledge-mcp.global.api.aws) connected directly (public endpoint, no auth)
3. Browser tools -- PlaywrightBrowser wrapped as LangChain tools (navigate, get_text, click, screenshot)
4. Code interpreter tools -- CodeInterpreter wrapped as LangChain tools (execute_code, execute_command)
5. Memory tool -- list_events using ListEventsCommand from the AWS SDK
6. Memory saving -- automatic CreateEventCommand after each response

All tools are merged and passed to createAgent, which lets Claude decide which tool to use.

Auto-build

With no Dockerfile present, the framework uses Heroku buildpacks to build the image. It reads package.json and uses the scripts.start field (node agents/index.js) as the entry point.

Requirements for auto-build:

• package.json with a start script
• A lockfile (package-lock.json)

Files

Cleanup

serverless remove

Related Examples

• LangGraph Gateway -- Gateway tools only
• LangGraph Browser -- Browser only
• LangGraph Code Interpreter -- Code interpreter only
• LangGraph Memory -- Memory only
• LangGraph Multi-Gateway -- Multiple gateways with different authorization