AWS launches Web Answer: one call replaces a citation-backed search pipeline for $10 per 1,000 requests

Anyone who has built an agent with internet access knows the routine: call a search API, decide which results to trust, extract and clean the text from pages, ask a model to summarize it, and then work out how to map every claim back to a source. Every step is code you have to maintain, and every step can break. AWS proposes replacing all of it with a single call.

What exactly has been released

Web Answer is now available in Amazon Bedrock AgentCore as a managed tool. An agent sends a question to AgentCore Gateway, which authenticates the request through IAM or JWT and calls Web Answer. Behind the scenes, the service plans multiple searches, runs them in parallel, reads key pages to verify facts, and searches again if the evidence is insufficient. The output is text in which every claim is linked to a source. The model is given the current date so it can look for up-to-date information.

For the agent, it is a standard MCP tool: the agent discovers it through tools/list and invokes it through tools/call. It works with Strands, LangChain, LangGraph, CrewAI, and any MCP client.

What it looks like in practice

Imagine a compliance bot at a bank. It is asked about Basel III minimum requirements for Tier 1 capital. Previously, you would have built the pipeline yourself. Now the agent calls web-answer-tool, and a few seconds later it receives an answer such as “at least 4.5% of risk-weighted assets,” with a link to bis.org. The response includes both markdown with [1] markers and a structured field containing an array of citations, making the links easy to display in an interface or retain for an audit.

Setup takes three steps. First, grant the Gateway service role the bedrock-agentcore:InvokeWebAnswer permission. Next, create a target with the connectorId web-answer. Finally, connect the agent. No separate keys are required: Gateway accesses the backend under its own role.

Source controls are particularly useful. A Gateway administrator can define domain allowlists and blocklists at the target level. The agent can narrow this filter for each call and can also restrict publication dates and country, but it cannot broaden the filter. In regulated industries, this is an important safeguard against the model unexpectedly citing a questionable website.

Web Answer or Web Search

The dividing line is who runs the model. Web Search returns ranked links, titles, and snippets; you perform synthesis with your own model and decide how to process the results. Web Answer handles the model, search loop, and attribution. AWS itself recommends combining the two: use Web Search for broad research that requires raw data, and Web Answer for focused factual questions within a larger workflow.

Limitations and what to watch

First, the price: $10 per 1,000 calls, or one cent per question. For an agent that asks dozens of questions per session, that adds up quickly, so repeated queries should be cached. The default limit is 10 requests per second per account, with increases available through support.

Only one region is currently available, us-east-1, with others promised later. If you have strict requirements to keep data in Europe, you will have to wait. Web Answer searches an Amazon-managed index and its knowledge graph without sending queries to third-party search engines. That is a privacy advantage, but you do not control the index's contents or ranking.

Most importantly, “an answer with a citation” is not the same as “a correct answer.” A link confirms where a claim came from, but it does not guarantee that the source is accurate or that the model interpreted it correctly. For legal and medical use cases, citations still need to be spot-checked, and preferably the allowed domains should be restricted to authoritative sources.

What to do

If you already operate in the AWS ecosystem and maintain a custom web-search pipeline, compare its maintenance cost with the one-cent-per-call price. For prototypes and supporting questions inside an agent, Web Answer looks like a sensible replacement. If you need full control over ranking, prompts, and the model, stay with Web Search or your own stack.