bedrock-agentcore（基岩 Agent 核心）

Amazon Bedrock AgentCore

Overview

Amazon Bedrock AgentCore is an agentic platform for building, deploying, and operating effective AI agents securely at scale—no infrastructure management needed. It provides framework-agnostic primitives that work with popular open-source frameworks (Strands, LangGraph, CrewAI, Autogen) and any model.

Purpose: Transform any AI agent into a production-ready application with enterprise-grade infrastructure

Pattern: Capabilities-based (6 independent service modules)

Key Principles (validated by AWS December 2025):

1. Framework Agnostic - Works with any agent framework or model
2. Zero Infrastructure - Fully managed, no ops overhead
3. Session Isolation - Complete data isolation between sessions
4. Enterprise Security - VPC, PrivateLink, identity integration
5. Composable Services - Use only what you need
6. Production Ready - Built for scale, reliability, and security

Quality Targets:

• Deployment: < 5 minutes from code to production
• Latency: Low-latency to 8-hour async workloads
• Observability: Full CloudWatch integration

When to Use

Use bedrock-agentcore when:

• Building production AI agents on AWS
• Need managed infrastructure for agent deployment
• Require session isolation and enterprise security
• Want to use existing agent frameworks (LangGraph, CrewAI, etc.)
• Need browser automation or code execution capabilities
• Integrating with existing identity providers

When NOT to Use:

• Simple Bedrock model invocations (use bedrock-runtime)
• Standard Bedrock Agents with action groups (use bedrock-agent)
• Non-AWS deployments

Prerequisites

Required

• AWS account with Bedrock access
• IAM permissions for AgentCore services
• Python 3.10+ (for SDK)

Recommended

• bedrock-agentcore-sdk-python installed
• bedrock-agentcore-starter-toolkit CLI
• Foundation model access enabled (Claude, etc.)

Installation

# Install SDK and CLI
pip install bedrock-agentcore strands-agents bedrock-agentcore-starter-toolkit

# Verify installation
agentcore --help

Core Services

1. AgentCore Runtime

Secure, session-isolated compute for running agent code.

Boto3 Client:

import boto3

# Data plane operations
client = boto3.client('bedrock-agentcore')

# Control plane operations
control = boto3.client('bedrock-agentcore-control')

Create Agent Runtime:

# Using starter toolkit
# agentcore configure -e main.py -n my-agent
# agentcore deploy

# Using boto3 control plane
response = control.create_agent_runtime(
    name='my-production-agent',
    description='Customer service agent',
    agentRuntimeArtifact={
        's3': {
            'uri': 's3://my-bucket/agent-package.zip'
        }
    },
    roleArn='arn:aws:iam::123456789012:role/AgentCoreExecutionRole',
    pythonRuntime='PYTHON_3_13',
    entryPoint=['main.py']
)
agent_runtime_arn = response['agentRuntimeArn']

Invoke Agent:

# Invoke deployed agent
response = client.invoke_agent_runtime(
    agentRuntimeArn='arn:aws:bedrock-agentcore:us-east-1:123456789012:agent-runtime/xxx',
    runtimeSessionId='session-123',
    payload={
        'prompt': 'What is my order status?',
        'context': {'user_id': 'user-456'}
    }
)

result = response['payload']
print(result)

Agent Entry Point Structure:

from bedrock_agentcore import BedrockAgentCoreApp
from strands import Agent

app = BedrockAgentCoreApp(debug=True)
agent = Agent()

@app.entrypoint
def invoke(payload):
    """Main agent entry point"""
    user_message = payload.get("prompt", "Hello!")
    app.logger.info(f"Processing: {user_message}")

    result = agent(user_message)
    return {"result": result.message}

if __name__ == "__main__":
    app.run()

2. AgentCore Gateway

Transforms existing APIs and Lambda functions into agent-compatible tools with semantic search discovery.

Create Gateway:

response = control.create_gateway(
    name='customer-service-gateway',
    description='Gateway for customer service tools',
    protocolType='REST'
)
gateway_arn = response['gatewayArn']

Add Gateway Target (Tool):

# Add an existing Lambda as a tool
response = control.create_gateway_target(
    gatewayId='gateway-xxx',
    name='GetOrderStatus',
    description='Retrieves order status by order ID',
    targetConfiguration={
        'lambdaTarget': {
            'lambdaArn': 'arn:aws:lambda:us-east-1:123456789012:function:GetOrder'
        }
    },
    toolSchema={
        'name': 'get_order_status',
        'description': 'Get the current status of a customer order',
        'inputSchema': {
            'type': 'object',
            'properties': {
                'order_id': {
                    'type': 'string',
                    'description': 'The unique order identifier'
                }
            },
            'required': ['order_id']
        }
    }
)

Synchronize Tools:

# Sync gateway tools for discovery
control.synchronize_gateway_targets(
    gatewayId='gateway-xxx'
)

3. Browser Runtime

Execute complex web-based workflows securely.

Start Browser Session:

response = client.start_browser_session(
    browserId='browser-xxx',
    sessionConfiguration={
        'timeout': 300,
        'viewport': {'width': 1920, 'height': 1080}
    }
)
session_id = response['browserSessionId']

Execute Browser Action:

# Navigate and interact
response = client.update_browser_stream(
    browserSessionId=session_id,
    action={
        'navigate': {'url': 'https://example.com'},
        'click': {'selector': '#submit-button'},
        'type': {'selector': '#search', 'text': 'query'}
    }
)

4. Code Interpreter

Safely execute code for tasks like data analysis and visualization.

Start Code Interpreter Session:

response = client.start_code_interpreter_session(
    codeInterpreterId='interpreter-xxx'
)
session_id = response['codeInterpreterSessionId']

Execute Code:

response = client.invoke_code_interpreter(
    codeInterpreterSessionId=session_id,
    code='''
import pandas as pd
import matplotlib.pyplot as plt

# Analyze data
df = pd.DataFrame({'x': [1,2,3,4,5], 'y': [2,4,6,8,10]})
plt.plot(df['x'], df['y'])
plt.savefig('output.png')
print(df.describe())
''',
    language='PYTHON'
)

output = response['output']
files = response['files']  # Generated files

5. Identity Integration

Native integration with existing identity providers for authentication and permission delegation.

Create OAuth2 Provider:

response = control.create_oauth2_credential_provider(
    name='okta-provider',
    credentialProviderVendor='OKTA',
    oauth2ProviderConfig={
        'clientId': 'your-client-id',
        'clientSecret': 'your-client-secret',
        'authorizationServerUrl': 'https://your-domain.okta.com/oauth2/default',
        'scopes': ['openid', 'profile', 'email']
    }
)

Create Workload Identity:

response = control.create_workload_identity(
    name='agent-identity',
    allowedRoleArns=['arn:aws:iam::123456789012:role/AgentRole']
)

Get Access Token:

# Get token for workload
response = client.get_workload_access_token(
    workloadIdentityId='identity-xxx'
)
access_token = response['accessToken']

6. Observability

Real-time visibility via CloudWatch and OpenTelemetry.

Enable Observability:

# In your agent entry point
from opentelemetry import trace
from opentelemetry.sdk.trace import TracerProvider

# Configure tracing
provider = TracerProvider()
trace.set_tracer_provider(provider)

# Entry point with OTel
# entryPoint: ['opentelemetry-instrument', 'main.py']

CloudWatch Metrics:

• Token usage per session
• Latency (p50, p95, p99)
• Session duration
• Error rates
• Tool call success/failure

Boto3 Client Reference

Data Plane (bedrock-agentcore)

Control Plane (bedrock-agentcore-control)

Quick Start: Hello World Agent

Step 1: Create Agent File

# main.py
from bedrock_agentcore import BedrockAgentCoreApp
from strands import Agent

app = BedrockAgentCoreApp()
agent = Agent(model="anthropic.claude-sonnet-4-20250514-v1:0")

@app.entrypoint
def invoke(payload):
    prompt = payload.get("prompt", "Hello!")
    result = agent(prompt)
    return {"response": result.message}

if __name__ == "__main__":
    app.run()

Step 2: Configure and Deploy

# Configure
agentcore configure -e main.py -n hello-world-agent

# Test locally
python main.py &
curl -X POST http://localhost:8080/invocations \
  -H "Content-Type: application/json" \
  -d '{"prompt": "Hello!"}'

# Deploy to AWS
agentcore deploy

# Test deployed
agentcore invoke '{"prompt": "Hello from production!"}'

Step 3: Invoke Programmatically

import boto3

client = boto3.client('bedrock-agentcore')

response = client.invoke_agent_runtime(
    agentRuntimeArn='arn:aws:bedrock-agentcore:us-east-1:123456789012:agent-runtime/hello-world',
    runtimeSessionId='test-session-1',
    payload={'prompt': 'What can you help me with?'}
)

print(response['payload'])

Error Handling

from botocore.exceptions import ClientError

try:
    response = client.invoke_agent_runtime(
        agentRuntimeArn=agent_arn,
        runtimeSessionId='session-1',
        payload={'prompt': 'test'}
    )
except ClientError as e:
    error_code = e.response['Error']['Code']

    if error_code == 'ResourceNotFoundException':
        print("Agent runtime not found")
    elif error_code == 'ValidationException':
        print("Invalid request parameters")
    elif error_code == 'ThrottlingException':
        print("Rate limited - implement backoff")
    elif error_code == 'AccessDeniedException':
        print("Check IAM permissions")
    else:
        raise

IAM Permissions

Minimum Execution Role

{
  "Version": "2012-10-17",
  "Statement": [
    {
      "Effect": "Allow",
      "Action": [
        "bedrock-agentcore:InvokeAgentRuntime",
        "bedrock-agentcore:StartBrowserSession",
        "bedrock-agentcore:InvokeCodeInterpreter"
      ],
      "Resource": "*"
    },
    {
      "Effect": "Allow",
      "Action": [
        "bedrock:InvokeModel",
        "bedrock:InvokeModelWithResponseStream"
      ],
      "Resource": "arn:aws:bedrock:*::foundation-model/*"
    },
    {
      "Effect": "Allow",
      "Action": [
        "logs:CreateLogGroup",
        "logs:CreateLogStream",
        "logs:PutLogEvents"
      ],
      "Resource": "arn:aws:logs:*:*:*"
    }
  ]
}

Related Skills

• bedrock-agentcore-policy: Cedar policy authoring and enforcement
• bedrock-agentcore-evaluations: Agent testing and quality evaluation
• bedrock-agentcore-memory: Episodic and short-term memory management
• bedrock-agentcore-deployment: Production deployment patterns
• bedrock-agentcore-multi-agent: Multi-agent orchestration (A2A protocol)
• boto3-eks: For EKS-hosted agents
• terraform-aws: Infrastructure as code

References

• references/gateway-configuration.md - Detailed gateway setup
• references/identity-integration.md - OAuth and workload identity
• references/troubleshooting.md - Common issues and solutions

Sources

• Amazon Bedrock AgentCore
• Boto3 BedrockAgentCore
• AgentCore SDK Python
• AgentCore Starter Toolkit