Amazon Bedrock AgentCore Web Search now generally available

Amazon Bedrock AgentCore Web Search is now generally available, giving agents a managed, MCP-compatible web search connector that draws on an Amazon-run index spanning tens of billions of documents and is refreshed continually, reflecting new content within minutes. The service is exposed as a managed target or connector you add to an AgentCore Gateway and is billed at $7 per 1,000 queries.

What is Web Search on Amazon Bedrock AgentCore?

Web Search on Amazon Bedrock AgentCore is a fully managed web-search capability that agents can discover and call via the Model Context Protocol (MCP). It is backed by a purpose-built web index that Amazon operates directly, spanning tens of billions of documents, and includes a built-in knowledge graph plus semantic snippet extraction optimized for model context.

The tool returns structured observations for each hit (title, url, publishedDate, text) and can return knowledge-graph observations with structured key/value facts. Amazon says the index is refreshed continually so new content can appear within minutes, and retrieval can combine the knowledge graph with semantic snippets so models receive the most relevant passages rather than raw HTML.

How do you wire it into an agent?

Attach the Web Search Tool as a target on an AgentCore Gateway using connectorId: "web-search" and let the Gateway handle authentication, schema, and endpoint resolution. The Gateway snapshots the tool schema, provisions the integration, and you discover the tool via tools/list from any MCP-compatible framework.

Practically, you add the target with create_gateway_target, configure outbound auth via credentialProviderConfigurations (examples use "GATEWAY_IAM_ROLE"), and let the Gateway assume an IAM service role that has permission to call the Web Search backend. The example permissions policy in the documentation shows two actions you must allow: bedrock-agentcore:InvokeGateway on your Gateway resource ARN and bedrock-agentcore:InvokeWebSearch on the AWS-owned tool ARN arn:aws:bedrock-agentcore:us-east-1:aws:tool/web-search.v1. Inbound authentication to the Gateway remains separate and is typically handled with OAuth or a JWT authorizer such as Amazon Cognito.

MCP-compatible frameworks such as Strands, LangChain, LangGraph, and CrewAI can discover and invoke the tool automatically. The tool returns results in the standard MCP tools/call envelope: a single content block of type text that contains a serialized JSON document with an id plus a results array of observations. Example observation fields include publishedDate (for example, "04:43AM, Wednesday, June 17 2026, PDT"), text, title, and url.

Why it matters

Grounding agents in the live web fixes the stale-knowledge problem without forcing teams to build and maintain search APIs, snippet extraction, credential management, or freshness pipelines. The managed index and knowledge graph aim to reduce factual drift by offering high-confidence entity facts and semantically relevant passages tuned for model context. For organizations with data-residency or egress concerns, the Gateway authenticates to an AWS-owned connector and routes queries entirely inside AWS, so customer queries are not sent to third-party search engines.

Operationally, that removes multiple engineering projects: procuring and managing third-party search keys and quotas, parsing inconsistent result formats, building snippet extraction logic, and owning freshness and coverage. The documentation also positions the Web Search Tool as a complement to Amazon Bedrock Knowledge Bases: use knowledge bases for private enterprise data and Web Search for public, rapidly changing information.

What to watch

Monitor how teams combine the Web Search Tool with model choices and agent frameworks. Key signals to watch are adoption of the connectorId "web-search" in production Gateways, the mix of public web search calls versus Bedrock Knowledge Base queries in hybrid agents, and cost patterns given the stated price of $7 per 1,000 queries. Also watch for published guidance or limits on the index coverage and the knowledge-graph scope to understand where the service provides high-confidence facts versus snippet-based grounding.