[アップデート] Amazon Bedrock AgentCore Managed Harness がプレビューリリースされました！

npm install -g @aws/agentcore@preview
agentcore --version
# 1.0.0-preview.1

agentcore create

What would you like to build?

❯ Harness (recommended) - Managed config-based agent loop, no code required
  Agent - Start with a template or bring your own code hosted on AgentCore Runtime
  Skip - I'll add resources later

Select model provider
Choose where to run your models

❯ Amazon Bedrock - Default: global.anthropic.claude-sonnet-4-6
  OpenAI - Default: gpt-5 (requires API key ARN)
  Google Gemini - Default: gemini-2.5-flash (requires API key ARN)

Custom environment
Optionally provide a custom container image for the harness runtime

❯ Default Environment - Includes Python, Bash, File tools
  Container URI - Use a pre-built container image (ECR URI)
  Dockerfile - Bring your own Dockerfile

Memory
Persistent memory lets the harness remember context across sessions

❯ No persistent memory - Harness does not retain context across sessions
  Enabled - Create persistent memory for this harness

Advanced settings (optional)
Configure tools, network, lifecycle, execution limits, truncation, or
session storage

❯ [✓] Tools - Add browser, code interpreter, MCP, or gateway tools
  [ ] Authentication - Inbound auth: AWS_IAM or Custom JWT
  [ ] Network - Deploy inside a VPC with custom subnets and security groups
  [ ] Lifecycle - Set idle timeout and max session lifetime
  [ ] Execution limits - Cap iterations, tokens, and per-turn timeout
  [ ] Truncation - Choose how context is managed when it exceeds limits
  [ ] Session Storage - Mount persistent storage for session data

Select tools for your harness
Choose built-in tools, MCP servers, or gateways

❯ [ ] AgentCore Browser - Web browsing and automation
  [✓] AgentCore Code Interpreter - Sandboxed code execution
  [ ] AgentCore Gateway - Connect via gateway
  [ ] Remote MCP Server - Connect to an MCP server

Review Configuration

  Name: MyHarness
  Model Provider: bedrock
  Model ID: global.anthropic.claude-sonnet-4-6
  Memory: Disabled
  Tools: AgentCore Code Interpreter

harnessSample/
├── agentcore/
│   ├── agentcore.json     # プロジェクト全体のスペック
│   ├── aws-targets.json   # デプロイ先リージョン/アカウント
│   └── cdk/               # CDK一式（L3はここに居る）
└── app/
    └── MyHarness/
        ├── harness.json       # ハーネス設定
        └── system-prompt.md   # システムプロンプト（別ファイル化）

{
  "name": "MyHarness",
  "model": {
    "provider": "bedrock",
    "modelId": "global.anthropic.claude-sonnet-4-6"
  },
  "tools": [
    {
      "type": "agentcore_code_interpreter",
      "name": "code-interpreter"
    }
  ],
  "skills": []
}

{
  "name": "harnessSample",
  "managedBy": "CDK",
  "harnesses": [
    { "name": "MyHarness", "path": "app/MyHarness" }
  ]
}

agentcore deploy

AgentCore Deploy

Project: harnessSample
Target: us-east-1:123456789012

[done]    Validate project
[done]    Check dependencies
[done]    Build CDK project
[done]    Synthesize CloudFormation
[done]    Check stack status
[done]    Publish assets
[done]    Deploy harnesses

Deployed 1 stack(s): AgentCore-harnessSample-default

agentcore invoke --harness MyHarness \
  --session-id "$(uuidgen)" \
  "121 + 23133 をCode Interpreterで計算して"

もちろんです！Code Interpreterで計算します。
🔧 Tool: code_interpreter

⚡ 4938 in · 121 out · 2.6s
🔗 Session: XXXXXXXX-XXXX-XXXX-XXXX-XXXXXXXXXXXX (use --session-id to continue)
  ✓
計算結果はこちらです！

**121 + 23133 = 23254**

⚡ 5073 in · 27 out · 2.6s
🔗 Session: XXXXXXXX-XXXX-XXXX-XXXX-XXXXXXXXXXXX (use --session-id to continue)

キザなセリフで返して。

agentcore deploy

ふっ…そんな計算、造作もないことだ。このオレ様の力を見せてやろう。
🔧 Tool: code_interpreter

  ✓
フハハ…結果は **23254** だ。

たかが足し算ごときに、このオレ様のCode Interpreterを使わせるとは…まったく、君も大した度胸をしている。

agentcore invoke --harness MyHarness \
  --model-id us.anthropic.claude-haiku-4-5-20251001-v1:0 \
  "121 + 23133 をCode Interpreterで計算して"

ふむ、xxを立てながら申し上げるならば、その計算程度は朝飯前。さっさと片付けてやろう。
🔧 Tool: code_interpreter

  ✓
ほうほう、期待通りじゃないか。**121 + 23133 = 23254** だ。

agentcore add harness --name invokeSample --with-invoke-script

Added harness 'invokeSample'.

app/invokeSample/
├── harness.json
├── invoke.py
└── system-prompt.md

{
  "name": "invokeSample",
  "model": {
    "provider": "bedrock",
    "modelId": "global.anthropic.claude-sonnet-4-6"
  },
  "tools": [],
  "skills": [],
  "memory": {
    "name": "invokeSampleMemory"
  }
}

import argparse
import json
import os
import sys
import uuid

import boto3

HARNESS_ARN = os.environ.get("HARNESS_ARN", "<your-harness-arn>")
REGION = os.environ.get("AWS_REGION", "<your-region>")
SESSION_ID = os.environ.get("SESSION_ID", str(uuid.uuid4()))

parser = argparse.ArgumentParser(description="Invoke an AgentCore Harness")
parser.add_argument("prompt", nargs="?", default="Hello!")
parser.add_argument("--raw-events", action="store_true")
parser.add_argument("--session-id", default=SESSION_ID)
args = parser.parse_args()

client = boto3.client("bedrock-agentcore", region_name=REGION)

response = client.invoke_harness(
    harnessArn=HARNESS_ARN,
    runtimeSessionId=args.session_id,
    messages=[
        {"role": "user", "content": [{"text": args.prompt}]}
    ],
)

for event in response["stream"]:
    if args.raw_events:
        print(json.dumps(event, default=str))
    else:
        if "contentBlockStart" in event:
            start = event["contentBlockStart"].get("start", {})
            if "toolUse" in start:
                tool = start["toolUse"]
                print(f"\n🔧 Tool: {tool.get('name', 'unknown')}", flush=True)
        elif "contentBlockDelta" in event:
            delta = event["contentBlockDelta"].get("delta", {})
            if "text" in delta:
                print(delta["text"], end="", flush=True)
        elif "messageStop" in event:
            stop_reason = event["messageStop"].get("stopReason", "")
            if stop_reason == "end_turn":
                print()
        elif "metadata" in event:
            usage = event["metadata"].get("usage", {})
            metrics = event["metadata"].get("metrics", {})
            latency = metrics.get("latencyMs", 0) / 1000
            print(
                f"\n⚡ {usage.get('inputTokens', 0)} in · "
                f"{usage.get('outputTokens', 0)} out · "
                f"{latency:.1f}s",
                file=sys.stderr,
            )

export HARNESS_ARN="arn:aws:bedrock-agentcore:us-east-1:123456789012:harness/harnessSample_MyHarness-XXXXXXXXXX"
export AWS_REGION="us-east-1"
python3 app/invokeSample/invoke.py "121 + 23133 をCode Interpreterで計算して"

ふっ…そんな計算、このオレに頼むとはなかなか見る目があるじゃないか。では、その程度の計算、華麗に片付けてやろう。
🔧 Tool: code_interpreter

⚡ 4943 in · 161 out · 4.1s
ふん…予想通りだ。結果を見ろ——

**121 + 23133 = 23254**

どうだ？この滑らかさ、まるでオレ自身のようだろう。次に難題があるなら、また頼ってくるがいい。オレはいつでもここにいる。✨
⚡ 5141 in · 98 out · 3.6s

🔗 Session: XXXXXXXX-XXXX-XXXX-XXXX-XXXXXXXXXXXX

python3 app/invokeSample/invoke.py --raw-events "121 + 23133 をCode Interpreterで計算して"

{"messageStart": {"role": "assistant"}}
{"contentBlockDelta": {"contentBlockIndex": 0, "delta": {"text": "\u304b"}}}
{"contentBlockDelta": {"contentBlockIndex": 0, "delta": {"text": "\u3057\u3053"}}}
...
{"contentBlockStop": {"contentBlockIndex": 0}}
{"contentBlockStart": {"contentBlockIndex": 1, "start": {"toolUse": {"toolUseId": "tooluse_XXXXXXXXXXXXXXXXXXXXXX", "name": "code_interpreter", "type": "tool_use"}}}}
{"contentBlockDelta": {"contentBlockIndex": 1, "delta": {"toolUse": {"input": "{\"code_in"}}}}
{"contentBlockDelta": {"contentBlockIndex": 1, "delta": {"toolUse": {"input": "terpreter_input\":{\"action\":{\"type\":\"executeCode\",\"code\":\"print(121 + 23133)\",\"language\":\"python\"}}"}}}}
{"contentBlockStop": {"contentBlockIndex": 1}}
{"contentBlockStart": {"contentBlockIndex": 2, "start": {"toolResult": {"toolUseId": "tooluse_XXXXXXXXXXXXXXXXXXXXXX"}}}}
...
{"messageStop": {"stopReason": "end_turn"}}
{"metadata": {"usage": {"inputTokens": 5096, "outputTokens": 96, "totalTokens": 5192}, "metrics": {"latencyMs": 3295}}}

import {
  AgentCoreApplication,
  AgentCoreMcp,
  type AgentCoreProjectSpec,
  type AgentCoreMcpSpec,
} from '@aws/agentcore-cdk';

export class AgentCoreStack extends Stack {
  public readonly application: AgentCoreApplication;

  constructor(scope: Construct, id: string, props: AgentCoreStackProps) {
    super(scope, id, props);
    const { spec, mcpSpec, credentials, harnesses } = props;

    this.application = new AgentCoreApplication(this, 'Application', {
      spec,
      harnesses,
    });

    if (mcpSpec?.agentCoreGateways && mcpSpec.agentCoreGateways.length > 0) {
      new AgentCoreMcp(this, 'Mcp', {
        projectName: spec.name,
        mcpSpec,
        agentCoreApplication: this.application,
        credentials,
        projectTags: spec.tags,
      });
    }
  }
}