AWS Bedrock AgentCore Web Search API: InvokeWebSearch gives AI agents web search in one call

What changed today#

Web Search on AgentCore is not new. AWS shipped it in June 2026 as a managed tool that agents reached through an AgentCore Gateway, over the Model Context Protocol. It worked, but it came with homework: create a gateway, add a web-search target, wire up the MCP plumbing, and pay the gateway charges on top of the $7 per 1,000 queries.

Tuesday's announcement removes the homework. InvokeWebSearch is a direct API: one call from the AWS SDKs or the AWS CLI, with nothing to provision first. AWS says the API is aimed at "applications, custom agent loops, and data pipelines" — in other words, the code you write yourself, not just agents running inside a framework.

What comes back is the same material the Gateway tool returned. Each result is ranked from Amazon's own web index — tens of billions of documents, refreshed continually, with a knowledge graph behind it, according to AWS's June launch post — and carries a text extract plus, when available, the page title, URL, and publication time. Every request can be filtered by domain and by publication time, so an agent can restrict itself to trusted sources or a recent window. AWS's own examples: finding recent product announcements, or checking a claim against trusted domains.

How it works in an agent loop#

Picture a research agent asked to verify a claim before it answers. It calls InvokeWebSearch with a natural-language query, gets back ranked snippets with their sources and dates, filters the request to a handful of trusted domains and the last 30 days, and grounds its answer in what came back. No scraper to maintain, no third-party search key to manage, no outbound credentials to rotate.

The enterprise controls are the point for AWS's core audience. Calls are authorized with IAM, so the same policies that govern the rest of a team's AWS estate govern search. The API can be called from inside a VPC through PrivateLink, keeping traffic off the public internet, and requests are logged as CloudTrail data events for audit. That is the whole pitch in one sentence: web search that behaves like any other AWS service.

The timing fits a broader Bedrock push. Kimi K3.1 Pro arrived on Bedrock last week, and Google Cloud answered with its own open AI-native stack the same morning. AWS is now filling in the plumbing around the models — the unglamorous pieces that decide whether an agent system is shippable.

What it costs — and the two other ways to buy the same search#

The price is the headline for anyone already paying for search: $7 per 1,000 queries, exactly what the Gateway route costs, minus the gateway. At that rate, a search-heavy agent making 10,000 queries a day spends $70 a day. But AWS now sells the same underlying search three different ways, and the differences matter.

The $12 figure belongs to the built-in Web Search tool AWS added to Bedrock model calls in August, which grounds a model's own response rather than serving a standalone search. The new API is the cheapest way to get Amazon's index into your own code — and the only one that needs no other AWS resource first. That matters in a week when the Open Markets Institute warned that whoever controls agent gateways controls the agent economy: a direct API is one less gate to pass through.

The catch#

Three limits are worth knowing before you build on it. First, geography: six regions at launch, and the list skips Frankfurt, Mumbai, Singapore, and GovCloud, so teams standardized on those regions wait. Second, depth: results are text extracts, not the full page. An agent that needs the whole document — a contract, a paper, a long report — still has to fetch it itself.

Third, it is an index, not the live web. AWS says its index reflects new content within minutes, and the June launch post describes tens of billions of documents refreshed continually. But a curated index is still a snapshot with a refresh lag, and AWS publishes no freshness guarantee for the newest pages. For grounding answers in current events, that is usually fine; for catching something published five minutes ago, verify before you trust it.