If you have ever built the "answer a question from the web" pipeline for an agent, you know the shape of it: search API, fetch pages, chunk, retrieve, synthesize, then bolt citations on at the end and hope they point at real sources. On 5 November 2026, AWS made that whole pipeline a single tool call — Web Answer is now generally available on Amazon Bedrock AgentCore.
One MCP tool call. A natural-language question goes in, a synthesized answer comes back, with citations attached. That is the whole pitch, and it is a genuinely useful one. But the price, the region limit and the rate limit decide whether it fits your agent, so this article will keep the pitch and the price tag separate.
What actually shipped
Web Answer is a managed connector in AgentCore Gateway. You point your agent's Gateway at it with the connectorId web-answer, and the Gateway's service role needs exactly one permission to call it: bedrock-agentcore:InvokeWebAnswer, against the AWS-owned ARN arn:aws:bedrock-agentcore:us-east-1:aws:tool/web-answer.v1.
That is the entire setup surface as AWS describes it. No search API key to manage, no retrieval pipeline to build, no citation formatting to write. Your agent asks a question in plain language and gets back a finished answer.
It is built on Web Search, the AgentCore tool that launched in June. The difference between the two is the difference between finding and answering — and it is worth being precise about, because it decides which one you should call.
Web Answer vs Web Search — which one you actually need
| Web Search (June) | Web Answer (GA, 5 Nov 2026) | |
|---|---|---|
| What it returns | Content it found — links and passages | A finished, synthesized answer |
| Citations | You build them | Included, inline [1] markers plus a citations list |
| Your pipeline | Search → retrieve → synthesize → cite | One tool call |
| Price | $7 per 1,000 queries | $10 per 1,000 tool calls |
| Use it when | You want raw material to reason over yourself | You want the answer, ready to show the user |
The honest read: Web Search is the tool you call when your agent needs to do its own thinking over fresh web content. Web Answer is the tool you call when the question has a factual answer and you want it synthesized and sourced without writing the pipeline. If your agent already has a retrieval step it trusts, Web Answer saves you nothing. If it does not — and most agents do not — Web Answer removes a whole subsystem.
What it costs — do the math before you wire it in
The pricing is simple, and that simplicity is the point: $10 per 1,000 tool calls, which is $0.01 per question answered. A thousand answers a day is $10 a day, $300 a month. Ten thousand answers a day is $100 a day.
There is also a rate limit: 10 transactions per second per account. At that ceiling you could theoretically make 36,000 calls an hour — which would cost $360 an hour — so the real limit on your usage is your budget, not the throughput. A chatty agent that calls Web Answer on every turn will feel that $0.01 per call quickly. An agent that calls it once per task will barely notice it.
The comparison that matters is against the $7 per 1,000 queries for Web Search. Web Answer costs about 43% more per call — and for that you get the synthesis and the citations done for you. If you were going to pay a model to synthesize search results anyway, the $3 difference per thousand is almost certainly cheaper than the tokens you would have spent. If you only needed links, Web Search is the cheaper call. Choose on what you do with the result, not on the headline price.
What the answer looks like
The response comes back in two parts. The content field is markdown with inline citations — the [1] markers sit in the answer text itself. The structuredContent.answer field carries the citations as a list, each with its source URL and domain, so your agent (or your UI) can render them properly instead of parsing them out of prose.
Two controls shape what comes back:
- Domain include/deny lists. You can restrict answers to domains you trust, or exclude domains you do not. For an agent that answers on behalf of a business, this is the difference between "the web said so" and "our approved sources said so."
- A publish-date filter. You can ask for recent information only, which is the right default for anything where yesterday's answer would be wrong.
These two controls are small features with outsized importance. A cited answer is only as trustworthy as the sources it cites, and a date filter is the only thing standing between a correct answer and a confidently outdated one. Use both from the start.
What to try first — if you try it at all
The lowest-risk experiment is the question your agent already gets asked most often and currently answers worst — the factual, time-sensitive question where it either hallucinates or gives a stale answer. Wire Web Answer to that one question type first, with a domain allow-list and a publish-date filter, and read the citations on the first fifty answers yourself.
Check three things in that trial: do the citations actually support the claims they are attached to, are the sources domains you would trust, and is the answer current? If all three hold, widen the rollout. If the citations are thin or the sources are odd, tighten the domain list before you blame the tool — the include/deny list is doing the quality work here, and it needs your judgment to work.
The experiment to hold off on is the "answer everything from the web" default. An agent that calls Web Answer on every turn is an agent spending $0.01 per turn for answers it may not need, and — more importantly — an agent whose answers you have not audited. Let the one-question trial prove the citation quality first.
The limits — stated plainly
- One region only, for now. Web Answer is available in us-east-1 (N. Virginia) only. If your agent stack lives elsewhere, that is a real constraint — your calls route to Virginia whether your users are there or not.
- 10 transactions per second per account. Fine for a single agent or a small fleet; a fleet of agents sharing one account will hit this ceiling fast. Plan your account layout around it.
- $0.01 per call adds up. A thousand answers a day is $300 a month. A chatty agent that answers from the web on every turn will turn a small feature into a visible line item.
- No accuracy numbers. AWS announced availability, not benchmarks. Nothing here tells you how often the citations support the claims, or how often the synthesized answer is wrong. The citation format is a trust mechanism, not a guarantee — read the sources, at least in the trial.
- Nothing here is hands-on. No Web Answer call was run for this article, and no citation quality, latency or cost was measured. Everything above comes from the AWS announcement and its getting-started post.
The bottom line is a practical one. Web Answer deletes a real pipeline — search, retrieve, synthesize, cite — and replaces it with one managed MCP call at a cent a question, with citations and domain controls built in. That is worth $10 per 1,000 calls if your agent needs fresh, factual answers and you do not already have a retrieval step you trust. It is not worth it if you only need links, if your stack cannot reach us-east-1, or if you were going to synthesize the results yourself anyway. Try it on your most-asked factual question, read fifty citations, then decide.
What is the one question your agent keeps answering badly that a cited web answer would fix?
Source: AWS Machine Learning Blog, "Getting started with Web Answer on Amazon Bedrock AgentCore" (5 November 2026, https://aws.amazon.com/blogs/machine-learning/getting-started-with-web-answer-on-amazon-bedrock-agentcore/). All facts come from the AWS announcement, checked 6 November 2026. No hands-on testing was performed for this article. This breaking article is not the daily buyer-intent post, and it contains no affiliate links.
