[アップデート]  Amazon Bedrock AgentCore Gateway のターゲットに Web Search ツールが追加されたので試してみた

[アップデート] Amazon Bedrock AgentCore Gateway のターゲットに Web Search ツールが追加されたので試してみた

AWS Summit New York で発表されたAmazon Bedrock AgentCore Gatewayのターゲットに追加されたWeb Searchを試してみました!
2026.06.18

はじめに

こんにちは、スーパーマーケットが好きなコンサル部の神野です。

2026年6月17日に開催された AWS Summit New York 2026 にて、Amazon Bedrock AgentCore の新機能が複数発表されました。その中のひとつが、今回紹介する Web Search ツールです!

これまで AI エージェントにウェブ検索機能を持たせるには、Tavily や Brave Search といった外部の検索 API を MCP サーバー経由で接続する必要がありました。別途サービスを利用する必要があり、AWSサービスで完結したら嬉しいのにな・・・と思っていたので待望のアップデートです。

AgentCore Gateway のコネクタターゲットとして組み込まれており、MCP 準拠のインターフェースでそのまま使えます。今回はコンソールから Gateway と Harness をセットアップして、実際に検索してみました!

Web Search on Amazon Bedrock AgentCore

公式の What's New から引用します。

Today, AWS announced the general availability of Web Search on Amazon Bedrock AgentCore, a fully managed tool that enables agents to ground responses in current, cited web knowledge with zero data egress from the customer's secured AWS environment.

https://aws.amazon.com/jp/about-aws/whats-new/2026/06/amazon-bedrock-agentcore-web-search/

ポイントをまとめると下記のとおりです。

  • フルマネージドなウェブ検索ツール。外部の検索 API を別途契約する必要がない
  • Amazon が独自に構築・運用するウェブインデックス(数百億ドキュメント規模)を使用。Alexa+ や Amazon Quick、Kiro などで培われた検索基盤がベース
  • ウェブインデックスに加えて、Amazon Knowledge Graph による構造化データも活用するマルチソースグラウンディング
  • クエリは AWS 内で完結し、外部の検索エンジンにデータが送信されない(ゼロデータエグレス)
  • MCP 準拠。Strands Agents、LangChain、CrewAI など任意の MCP クライアントから利用可能

良さげな感じですね!精度がどれぐらいなのか気になります!

詳細は公式ブログも併せてご確認ください。

https://aws.amazon.com/blogs/aws/announcing-web-search-on-amazon-bedrock-agentcore-ground-your-ai-agents-in-current-accurate-web-knowledge/

前提

  • リージョン: US East (N. Virginia) us-east-1 のみ対応(2026年6月時点)

    • 現時点では東京リージョン対応していないので注意が必要です。
  • 料金: Web Search ツール自体の追加料金はなし。Gateway のデータ転送料金のみ発生

    • ツール自体の追加料金がかからないのは嬉しいです!

やってみた

コンソールを中心に早速試してみます!

Gateway の作成

まず AgentCore の Gateway を作成します。左メニューの Gateways から「Create Gateway」を押下します。

01-gateway-list

Define gateway details

Gateway 名を入力し、IAM permissions は「Create default role」を選択します。必要なサービスロールが自動で作成されます。

02-gateway-details

Configure Inbound Identity

Inbound Auth type は「Use IAM permissions」を選択します。 Harness から IAM で認証・認可が行われるため、追加の設定は不要です。

03-inbound-identity

Add targets

ここが今回新しく追加された設定となります。ターゲットの設定で下記を選択します。

  • Select a target protocol: MCP target
  • Pass-through: Do not use pass-through
  • Tool type: Connectors
    • 新規追加!

Pre-configured targets のセクションに「Web Search tool - new」が表示されているので、これを選択します。

04-add-targets

Review and create

設定内容を確認して「Create Gateway」を押下します。Target type が webSearch になっていれば OK です。

05-review-create

Harness の作成

Gateway ができたので、次に Harness を作成して Playground から検索を試します。

左メニューの Harness から「Quick create Harness」のドロップダウンを開き、「Advanced create Harness」を選択します。

05a-harness-dropdown

Harness の設定

Harness 名を入力し、Tools セクションで Gateway を有効にします。先ほど作成した Gateway を選択し、Outbound Auth configuration は IAM role を指定します。

06-create-harness

作成が完了すると、Harness の詳細画面が表示されます。Status が Ready になっていれば準備完了です。

07-harness-detail

Harness Playground で試す

右上の「Test Harness」から Playground を開きます。モデルは Claude Sonnet 4.6 を使用しました。

07a-harness-test

クラスメソッド株式会社について聞いてみる

まずは「クラスメソッド株式会社について教えて」と聞いてみました。

08-harness-playground-1

09-harness-playground-2

基本情報、事業内容、受賞実績、グローバル展開などがまとめられています! ソースとなる検索結果のスニペットをもとに、構造化された回答が生成されています。

ただし、公開されている情報と照らし合わせると気になる点もありました。例えば従業員数が「約600名」と回答されていますが、公式サイトでは約820名(2025年7月時点)です。ミッションの日本語訳も現在の正式な表現とは異なっていました。検索結果に含まれる古い情報やサードパーティの情報がそのまま採用されていました。

下部のツールコールのトレースを見ると、WebSearch ツールが複数回呼び出されている様子が確認できます。

10-tool-calls

クエリの内容を見ると「クラスメソッド株式会社 Classmethod Inc 事業...」「classmethod.jp 会社情報 設立 従業員数」のように、エージェントが自律的にクエリを工夫して複数回検索しています。ただ、検索結果には株式会社楽生やクロスギルドなど関連性の薄い日本語サイトも混ざっていました。

自分の最近のブログ記事を聞いてみる

続いて「クラスメソッドの神野さんが一番直近で書いたブログ記事を教えて」と聞いてみました。

11-harness-playground-blog

直近の記事一覧はおおむね取れていますが、最新の記事はヒットしていませんでした。

ツールコールのトレースを見ると検索クエリが site:dev.classmethod.jp 神野 最新記...dev.classmethod.jp 神野 author D... のように工夫されているのですが、ヒットしている URL が https://dev.classmethod.jp/en/... と英語版のページばかりでした。DevelopersIO の日本語記事ページではなく、英語翻訳版のほうがインデックスされているようです。先ほどのクラスメソッドの検索結果と合わせて、現時点では英語コンテンツのインデックスが中心なのかなという印象を受けました。

英語で質問してみる

日本語クエリの精度が気になったので、逆に英語で質問してみます。「What new features were announced for Amazon Bedrock AgentCore at AWS Summit New York 2026?」と聞いてみました。

12-english-query

おおお、 AWS Summit NYC 2026 で発表された AgentCore 関連のアップデートがかなり網羅的にまとめられているように見えます!

  • Web Search on AgentCore(GA)
  • Amazon Bedrock Managed Knowledge Base(新規)
  • Optimization Capabilities(Insights: Preview / Batch Evals: GA)
  • AgentCore Harness(GA)
  • AWS Context(Coming Soon)

各項目が公式ソース付きで整理されていました。

回答の詳細
Here is a comprehensive summary of all the new features announced for **Amazon Bedrock AgentCore** at **AWS Summit New York 2026** (June 17, 2026):

---

## 🚀 Amazon Bedrock AgentCore — New Feature Announcements at AWS Summit New York 2026

The overarching theme was: **"Build, connect, and optimize production agents securely at scale."** The announcements fell into three pillars: broader knowledge, production observability/optimization, and scalable governance.

---

### 1. 🔍 Web Search on Amazon Bedrock AgentCore — **Generally Available**
- A **fully managed web search tool** that enables agents to ground their responses in **current, cited web knowledge**.
- Uses **AWS's own search index**, meaning queries never leave the customer's secured AWS environment — **zero data egress**.
- Available at **no additional cost** (customers pay only for Gateway data transfer charges).
- Now **GA in the US East (N. Virginia)** Region.
- *Source: [AWS News Blog](https://aws.amazon.com/blogs/aws/announcing-web-search-on-amazon-bedrock-agentcore-ground-your-ai-agents-in-current-accurate-web-knowledge/)*

---

### 2. 🗂️ Amazon Bedrock Managed Knowledge Base — **New Launch**
- A new capability that **automatically manages and scales the end-to-end RAG (Retrieval-Augmented Generation) pipeline** for enterprise knowledge agents.
- Features include:
  - **Native data connectors** selectable from a dropdown
  - **Smart Parsing** for automatic multi-format data preparation
  - **Agentic Retriever** for complex multi-step queries
  - **Auto-generated IAM roles** for permissions management
- **Natively integrated with AgentCore Gateway** as a pre-built target type (alongside MCP server, Lambda ARN, REST API, etc.), with built-in observability, policy enforcement, and automatic permission management.
- *Source: [AWS News Blog](https://aws.amazon.com/blogs/aws/introducing-amazon-bedrock-managed-knowledge-base-for-faster-more-accurate-enterprise-ai-applications/)*

---

### 3. 🛠️ New Optimization Capabilities — **Continuous Improvement for Production Agents**
- Turns **production traces into continuous improvement loops** for agents, addressing silent failures that produce no error signal.
- The loop covers: understanding what agents are doing → generating data-grounded fixes → proving they work.
- New capabilities include:
  - **Failure, intent, and trajectory insights** — available in **preview** in 13 AWS Regions
  - **Batch evaluations, recommendations, and A/B tests** — **generally available** in 14 AWS Regions
- *Source: [AWS What's New](https://aws.amazon.com/about-aws/whats-new/2026/06/amazon-bedrock-agentcore-new-optimization-capabilities/)*

---

### 4. ⚙️ AgentCore Harness — **Generally Available**
- The **managed agent harness** (previously in preview) is now **generally available**.
- Allows developers to **build and run production-grade AI agents in minutes — without coding orchestration loops** — simply by defining the agent's model, tools, skills, and instructions in configuration.

---

### 5. 🌐 Coming Soon: AWS Context (Integration with AgentCore)
- A brand-new companion service, **AWS Context**, was announced that automatically maps relationships across organizational data into a **knowledge graph**.
- It provides **agentic search** so agents can access governed data relationships, business rules, and domain knowledge at runtime — deeply integrated with the AgentCore ecosystem.

---

### Summary Table

| Feature | Status | Key Benefit |
|---|---|---|
| Web Search on AgentCore | **GA** | Current, cited web knowledge inside AWS environment |
| Managed Knowledge Base | **New Launch** | Fully managed RAG pipeline with native AgentCore Gateway integration |
| Optimization Capabilities (Failure/Intent/Trajectory Insights) | **Preview** (13 Regions) | Detect silent production failures |
| Optimization Capabilities (Batch Evals, A/B Tests) | **GA** (14 Regions) | Continuous agent improvement loop |
| AgentCore Harness | **GA** | No-code orchestration for production agents |
| AWS Context (Knowledge Graph) | **Coming Soon** | Org-wide agentic knowledge access |

These announcements were positioned as helping customers **connect agents to broader knowledge sources, continuously improve them in production, and enforce controls that scale as agents grow more capable**.

日本語を中心とした情報にアクセスした時よりは回答の正確性に差があるように感じました。現時点では、英語圏に情報がある方が精度が良さそうな印象です。

参考までに、AWS NYC Summit の情報は下記記事がわかりやすくまとまっています。この記事と検索結果を比較するとどれぐらいの精度かわかりやすいと思います。

https://qiita.com/hayao_k/items/44b25e2a51d12482a308

おわりに

外部 API なしでエージェントにウェブ検索を組み込めるのは、セットアップも含めて簡単でした!

日本語の検索精度や回答の正確性はまだ改善の余地がありそうですが、マネージドサービスとして提供されてAWS内で完結するのは嬉しいです。
早速取り入れてみて、実際にどう使って行くかは検証していきたいです。

本記事が少しでも参考になりましたら幸いです。最後までご覧いただきありがとうございましたーー!

補足

公式ドキュメントから、知っておくと便利な仕様をいくつかピックアップします。

https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/gateway-target-connector-web-search-tool.html

Input Schema

tools/call で呼び出す際のパラメータは下記のとおりです。

フィールド 必須 説明
query string Yes 検索クエリ。200文字以内
maxResults integer No 返却する最大件数。1〜25の範囲。デフォルトは10

Response Format

レスポンスは MCP 準拠の JSON 形式で返却されます。各検索結果には text / url / title / publishedDate が含まれます。

レスポンス例
{
  "isError": false,
  "content": [
    {
      "type": "text",
      "text": "{\"id\":\"824f89d0\",\"results\":[{\"text\":\"Python 3.13 was released on...\",\"publishedDate\":\"2024-10-07\",\"url\":\"https://example.com/python/releases/3.13\",\"title\":\"Python 3.13 Release Highlights\"}]}"
    }
  ]
}

Domain Filtering

ドメインの拒否リストを設定することで、特定のドメインを検索対象から除外できます。

Acceptable Use

検索結果を利用する際は、ソースの引用とリンクをエンドユーザーに表示する必要があります。また、検索結果のバルク抽出・保存や、競合するインデックス・データベースの構築に使用することは禁止されています。

この記事をシェアする

関連記事