amazon-bedrock-agentcore

name

amazon-bedrock-agentcore

description

Expert guide for designing, developing, testing, and deploying AI agent products using Amazon Bedrock AgentCore. Covers ALL AgentCore components (Runtime, Gateway/MCP, Memory, Identity, Browser, Code Interpreter, Observability, Evaluations, Policy, Registry, Payments) with CDK L2 constructs for IaC. ALWAYS invoke this skill when the user mentions any of the following: - "AgentCore", "bedrock-agentcore", "agent runtime", "MCP gateway", "agent memory" - Deploying an AI agent to AWS (serverless, containerized) - Converting APIs or Lambda functions to MCP tools - Agent identity / OAuth / credential management on AWS - Browser automation or code interpreter for AI agents - Evaluating agent quality (LLM-as-a-Judge, automated evals) - Cedar policy or governance for AI agents - Strands, LangGraph, CrewAI, LlamaIndex deployment to AWS - CDK infrastructure for AI agents (aws_bedrockagentcore, Runtime, Gateway constructs) - "エージェント", "MCP サーバー", "エージェント デプロイ" (Japanese equivalents)

Amazon Bedrock AgentCore — Expert Skill

Amazon Bedrock AgentCore is a framework-agnostic, model-agnostic platform for building, deploying, and operating AI agents at scale. It works with any open-source framework (Strands, LangGraph, CrewAI, LlamaIndex, OpenAI Agents SDK, Google ADK) and any foundation model.

Quick Decision: Which Component Do You Need?

User Goal	AgentCore Component	Reference
Deploy agent code serverlessly (any framework)	Runtime	runtime.md
Single-API agent loop (no container needed)	Harness	AWS docs
Expose REST API / Lambda as MCP tools	Gateway	gateway.md
Connect to existing MCP servers	Gateway (MCP target)	gateway.md
Multi-turn conversation memory	Memory	memory.md
Agent auth with Okta/Cognito/Entra	Identity	identity-policy.md
Policy/governance for tool calls	Policy (Cedar)	identity-policy.md
Agent executes Python/JS code in sandbox	Code Interpreter	tools.md
Agent navigates websites / fills forms	Browser	tools.md
Trace, debug, monitor agent in prod	Observability	observability-evals.md
Measure agent quality (LLM-as-a-Judge)	Evaluations	observability-evals.md
IaC with AWS CDK	CDK L2 constructs	cdk-patterns.md
Strands / LangGraph / CrewAI integration	Framework patterns	frameworks.md

Component Overview

┌────────────────────────────────────────────────────────────────────┐
│                   Amazon Bedrock AgentCore                          │
│                                                                      │
│  ┌─────────┐  ┌─────────┐  ┌────────┐  ┌──────────┐  ┌─────────┐ │
│  │ Runtime │  │ Gateway │  │ Memory │  │ Identity │  │  Policy │ │
│  │(deploy) │  │ (MCP)   │  │(state) │  │  (auth)  │  │ (Cedar) │ │
│  └─────────┘  └─────────┘  └────────┘  └──────────┘  └─────────┘ │
│                                                                      │
│  ┌──────────────┐  ┌─────────┐  ┌─────────────┐  ┌─────────────┐ │
│  │Code Interpret│  │ Browser │  │Observability│  │ Evaluations │ │
│  │  (sandbox)   │  │ (web)   │  │  (OTEL)     │  │(LLM judge)  │ │
│  └──────────────┘  └─────────┘  └─────────────┘  └─────────────┘ │
│                                                                      │
│  ┌──────────┐  ┌──────────┐  ┌─────────┐                          │
│  │ Registry │  │ Payments │  │ Harness │                          │
│  │(catalog) │  │  (x402)  │  │(managed)│                          │
│  └──────────┘  └──────────┘  └─────────┘                          │
└────────────────────────────────────────────────────────────────────┘

Standard Workflow

When the user asks to build or deploy something with AgentCore, follow this sequence:

1. UNDERSTAND scope → Which components? Which framework?
2. DESIGN architecture → Read relevant reference files
3. IMPLEMENT → Agent code + CDK IaC (cdk-patterns.md)
4. TEST locally → Python/SDK invocation tests
5. DEPLOY → CDK deploy
6. VERIFY → Observability + Evaluations

API Endpoints (Critical — Wrong Endpoint = Error)

Plane	Client Name	Use For
Control	bedrock-agentcore-control	Create/manage runtimes, gateways, memory, evaluations
Data	bedrock-agentcore	Invoke agent runtimes
Gateway Data	{gatewayId}.gateway.bedrock-agentcore.{region}.amazonaws.com	Call a specific gateway's MCP endpoint

Refer to AgentCore endpoints and quotas for current endpoints.

CDK Module (L2 Constructs)

import * as agentcore from 'aws-cdk-lib/aws_bedrockagentcore'; // TypeScript

from aws_cdk import aws_bedrockagentcore as agentcore  # Python

Available in CDK ≥ 2.221.0 (alpha — breaking changes possible).

Key constructs: Runtime, Gateway, Memory, BrowserCustom, CodeInterpreterCustom, OnlineEvaluation, WorkloadIdentity

See cdk-patterns.md for complete IaC examples.

Key Constraints (Always Apply)

• ARM64 only for Runtime containers — x86 images will not start
• Protocol must be chosen before building container (HTTP / MCP / A2A / AG-UI)
• Endpoint required — Runtime is not invocable until a Runtime Endpoint is created and READY
• No secrets in environment variables — use Secrets Manager
• Production Runtimes must have authorizer — never deploy unauthenticated
• CDK alpha status — aws_bedrockagentcore module may have breaking changes; pin CDK version

Pricing (Consumption-based, No Upfront)

Use Case	Estimated Cost
Simple agent (low volume)	~¥450/month
Customer support agent (high volume)	~¥11,000/month
Exact rates	AgentCore Pricing

When to Read Reference Files

• Building/deploying agent → runtime.md
• Exposing APIs as MCP tools → gateway.md
• Adding memory to agent → memory.md
• Auth / identity setup → identity-policy.md
• Browser / code sandbox → tools.md
• Monitoring / evaluation → observability-evals.md
• Full CDK stacks → cdk-patterns.md
• Framework integration (Strands/LangGraph) → frameworks.md