AWS launches a direct Web Search API for AI agents, without a gateway

What launched on November 11

AWS announced a standalone Web Search API on Amazon Bedrock AgentCore on November 11, 2026. Developers can now call InvokeWebSearch from an AWS SDK or the AWS CLI, using existing IAM credentials, without creating a gateway or search target first. AWS lists the service at $7 per 1,000 search queries, with no gateway charges on this direct route.

This is a new access path to an existing search service, not a new foundation model or a first launch of AgentCore Web Search. Previously, the documented route exposed search through AgentCore Gateway. The direct API matters for AI builders who already own the orchestration loop and want to add current web evidence without adopting that extra tool-serving layer.

The service returns ranked results from an Amazon-operated web index. Each result has a text extract; a title, URL and publication time are returned when available. It does not itself produce a finished answer or prove that an agent’s final response is correctly grounded. Retrieval and synthesis remain separate responsibilities.

Availability, identifiers and practical limits

The operation is InvokeWebSearch on the bedrock-agentcore data plane; in Boto3 it is client.invoke_web_search. Requests use webSearchId="web-search.v1" and searchQuery for the query text. The IAM action is bedrock-agentcore:InvokeWebSearch, scoped to the AWS-managed web search tool resource. AWS signs requests with Signature Version 4 rather than requiring a separate search API key.

The six documented Regions are us-east-1, us-east-2, us-west-2, eu-west-1, ap-northeast-1 and ap-southeast-2. The developer guide explicitly says the direct API is not available in AWS GovCloud (US-West), where readers should use the existing connector instead. Do not infer availability from the broader AgentCore footprint.

AWS allows searchQuery values up to 200 characters and maxResults from 1 to 25. Omitting maxResults returns up to 25 results; the console starts at 10, so copying console expectations into code can change the amount of retrieved context. The default quota is 10 requests per second per account per Region, with increases available by request.

AWS specifies Boto3 1.43.127 or later for its Python example. Framework support is documented in langchain-aws 1.9.3 and the AgentCore Python SDK 1.26.0 or later. These are vendor-stated integration requirements; we did not install those versions or make a paid search call.

What changes for custom agent loops

For an application that already implements tool execution, the new integration is a direct request followed by structured results. Application code can decide when to search, how to retry, which extracts to pass to a model and how to cache permitted output. AWS positions the API for custom orchestration, existing applications and research or fact-checking pipelines; its gateway remains the recommended path for many shared agent-tool deployments.

Domain filters can include or exclude up to 100 domains in each list, and a root domain also matches its subdomains. Publication bounds use publishedAfter and publishedBefore, with inclusive UTC date-times. Use an explicit time: the guide warns that an SDK or CLI date without a time becomes midnight at the start of that day, unlike the connector’s whole-day end-date behavior.

The important policy difference is who supplies those filters. A gateway target can impose administrator-defined domain lists on every request. The direct API carries filters in each request. Our recommendation is to construct required filters in trusted application code, rather than allowing a model to remove them while composing a tool call. A domain match is a source-selection aid, not evidence that the page’s claims are true.

Pricing, networking and the limits of the evidence

The launch post and current AgentCore pricing page agree on $7 per 1,000 queries. AWS says direct API calls meter to the same Web Search usage type as connector searches, but do not incur gateway charges. This is not a complete agent-task price: model inference, application execution, logs and optional PrivateLink interface endpoints are separate considerations. We have not measured an end-to-end bill.

AWS says queries are served inside AWS from an index Amazon operates and are not sent to a third-party search engine. The API supports PrivateLink for callers in private networks. Those are vendor-described properties; we did not trace network traffic, audit retention or test endpoint policies. Private connectivity should not be described as proof that all retrieved content or all subsequent model calls remain inside a private network.

CloudTrail records InvokeWebSearch as data events, which are not enabled by default. AWS says those events include the tool and maxResults, but not search text or filters. Teams needing a query-level provenance record must design appropriate application logging with their own data-handling rules; turning on CloudTrail alone will not capture a complete search transcript.

Neither reviewed source set establishes comparative relevance, index completeness, freshness, latency or answer accuracy. The announcement’s example search output is AWS’s example, not an AccessAllGPT test result. A lower integration burden is plausible; a better research answer remains something to evaluate.

What builders should do next

Start with one non-sensitive research question in a supported Region and an IAM role limited to the search action. Set maxResults explicitly, choose a small set of first-party domains and use full UTC timestamps for the publication window. Inspect returned URLs and dates rather than assuming every extract has complete metadata. Then open the original sources before accepting a consequential claim.

Retain the citations your application uses. The developer guide requires source citations and links to be retained and displayed in end-user output that uses search results. It also prohibits bulk extraction, storage or reproduction of result content, and using the service to populate a competing index or database. Treat those terms as a product-design constraint before proposing a large crawling or caching workflow.

Our decision: trial the API when your application already owns the agent loop and IAM integration, and the six-Region footprint fits. Retain the gateway path when shared tool discovery, JWT authorization or administrator-controlled target policies are the actual requirement. Do not migrate a functioning connector just because a lower-level API now exists.

Disclosures

AccessAllGPT received no payment, credits or private access for this report. AWS did not review or endorse it. AccessAllGPT Research is operated by NeuralArc, is independent, and is not affiliated with AWS. Publication-wide relationships appear on the disclosures page.