A minimal LangGraph agent demonstrating core AgentCore concepts using code deployment.

Alternative: See langgraph-basic-docker for the same agent using Docker/container deployment.

• BedrockAgentCoreApp: Integration pattern for AgentCore Runtime
• LangGraph: Agent orchestration with state management
• Claude Sonnet 5: High-performance reasoning model
• Simple Tools: Calculator and time tools
• Code Deployment: Python-based deployment (no Dockerfile needed)

User Request → AgentCore Runtime → agent_invocation()
                                      ↓
                                   LangGraph
                                      ↓
                               Claude Sonnet 5
                                      ↓
                          (uses built-in tools)
                                      ↓
                                   Response

• AWS account with Bedrock model access (Claude Sonnet 5)
• Enable access to the global.anthropic.claude-sonnet-5 inference profile in Bedrock console
• Serverless Framework v4+
• AWS credentials configured
• No Docker required (unlike container deployment)

Important: This example uses the global cross-region inference profile for better availability and throughput. Direct model IDs may not support on-demand invocation.

# From this directory
serverless deploy

The framework will:

• Package Python code with dependencies
• Upload to Amazon S3
• Deploy AgentCore Runtime with managed Python runtime
• Output the invocation URL

Using the provided test script:

# Update RUNTIME_ARN in test-invoke.py first
python3 test-invoke.py

Or invoke programmatically with boto3:

import boto3
import json
import uuid

client = boto3.client('bedrock-agentcore', region_name='us-east-1')

response = client.invoke_agent_runtime(
    agentRuntimeArn='YOUR_RUNTIME_ARN',  # From deploy output
    runtimeSessionId=str(uuid.uuid4()),
    payload=json.dumps({"prompt": "What is 25 multiplied by 4?"}).encode()
)

# Parse response
result = json.loads(response['response'].read())
print(result)

Important: You cannot invoke AgentCore runtimes directly via curl. You must use the AWS SDK with the bedrock-agentcore client and the invoke_agent_runtime API method.

Test locally before deploying:

serverless dev

agent.py implements a simple LangGraph agent:

1. Initialize LLM: Uses Claude Sonnet 5 via Bedrock Converse API
2. Define Tools: Adds calculator and time tools using @tool decorator
3. Build Graph: Creates a state machine with chatbot and tool nodes
4. Entrypoint: @app.entrypoint decorator marks the invocation function
5. Process Messages: Handles requests and returns responses

START → chatbot → [decide: use tool or respond]
           ↑            ↓
           └─── tools ←┘

• chatbot node: Invokes Claude with tool availability
• tools node: Executes tools if requested
• conditional edge: Decides whether to use tools based on LLM response

The configuration specifies:

• handler: agent.py - Entry point file (triggers code deployment mode)
• package.patterns - Files to include

Dependencies in requirements.txt are bundled automatically by AgentCore's code deployment packaging.

AgentCore automatically:

• Packages code with dependencies
• Uploads to S3
• Deploys to managed Python runtime

The default model is global.anthropic.claude-sonnet-5, read from the MODEL_ID environment variable in agent.py. Override it via the MODEL_ID env var:

ai:
  agents:
    chatbot:
      handler: agent.py
      environment:
        MODEL_ID: us.anthropic.claude-opus-4-1-20250805-v1:0 # Different model

Add tools in agent.py using the @tool decorator:

from langchain_core.tools import tool

@tool
def search_database(query: str) -> str:
    """Search the product database."""
    # Your database logic here
    return f"Found products matching: {query}"

# Add to tools list
tools = [get_current_time, add, multiply, search_database]

Add optional runtime configuration:

ai:
  agents:
    chatbot:
      handler: agent.py
      environment:
        CUSTOM_VAR: value
      lifecycle:
        idleRuntimeSessionTimeout: 900 # Idle timeout (60-28800 seconds)
        maxLifetime: 3600 # Max lifetime (60-28800 seconds)

Remove all resources:

serverless remove

• Add gateway tools - Expose Lambda functions as agent tools
• Add memory - Enable conversation persistence