What AWS released

On 10 November 2026, AWS announced general availability of the Web Search API on Amazon Bedrock AgentCore. Until now, AgentCore's web search was only reachable as an MCP tool behind an AgentCore Gateway. MCP is the Model Context Protocol, the open standard agents use to discover and call tools. The new InvokeWebSearch operation lets ordinary code send a query and get ranked results back in one call. You don't create a gateway, a target or any other resource first. AWS's launch post on the Machine Learning blog, published 11 November, names the main users as custom agent loops, LangChain or LlamaIndex apps, and non-agent pipelines such as fact-checking and monitoring. The same post still recommends the Gateway for most agentic workloads.

For New Zealand teams, the useful detail is that Asia Pacific (Sydney), ap-southeast-2, is one of the six regions where the API is available. The Auckland region, ap-southeast-6, opened in September 2025 but isn't on that list. Workloads running in Auckland will need to call the Sydney endpoint.

How the API works

A request carries webSearchId (always web-search.v1 for the AWS-managed tool) and a searchQuery of 1 to 200 characters. The request reference lists three optional controls: maxResults (1 to 25, defaulting to 25), a domain filter with include and exclude lists of up to 100 domains each, and a publication date filter using publishedAfter and publishedBefore. A root domain also matches its subdomains. The API rejects any unknown top-level field with a ValidationException.

The response is a single ranked results array with no pagination. Every result has a text extract. The title, url and publishedAt fields appear only when AWS knows them, so your code should handle a result that has no URL.

According to the Web Search Tool documentation, the index behind the tool is built and run by Amazon, spans tens of billions of documents and refreshes within minutes. It also returns extracted passages instead of raw HTML. AWS's blog states that "Queries are served inside AWS and aren't sent to a third-party search engine."

Setting it up

These steps follow the getting started guide.

  1. Update your SDK. The docs say older Boto3 and AWS CLI versions don't include the operation. AWS's blog gives the minimums as Boto3 1.43.127, langchain-aws 1.9.3 and the AgentCore Python SDK 1.26.0.
  2. Grant the IAM permission. The resource is an AWS-owned tool, so the account field in the ARN is the literal string aws. Replace <Region> with ap-southeast-2 for Sydney:
{
    "Version": "2012-10-17",
    "Statement": [
        {
            "Sid": "AllowInvokeWebSearch",
            "Effect": "Allow",
            "Action": "bedrock-agentcore:InvokeWebSearch",
            "Resource": "arn:aws:bedrock-agentcore:<Region>:aws:tool/web-search.v1"
        }
    ]
}
  1. Test from the CLI. This is the command from AWS's blog. Add --region ap-southeast-2 as the docs example does if your default region is elsewhere:
aws bedrock-agentcore invoke-web-search \
    --web-search-id web-search.v1 \
    --search-query "latest EKS features" \
    --max-results 5
  1. Call the API from Python. Create a bedrock-agentcore client with Boto3, call client.invoke_web_search(...) with the same fields, and loop over response["results"]. The docs example wraps the call in a ClientError handler.
  2. Handle errors. The troubleshooting page says to retry ThrottlingException (HTTP 429) and InternalServerException with exponential backoff. For ServiceUnavailableException, wait for retryAfterSeconds when the response includes it. An AccessDeniedException usually means the ARN has the wrong region, or your own account ID where aws should be.

Costs and quotas

The AgentCore pricing page lists Web Search at US$7 per 1,000 queries. The announcement confirms that's the same rate as searching through the Gateway, without gateway charges on top. The AgentCore release notes add a November 2026 entry making the first 5,000 queries each month free for every AWS account.

The free queries come with conditions set out in the documentation. The API and the Gateway tool draw on the same allowance, and so does every account in an AWS Organization. The allowance resets at the start of each calendar month in UTC. Once it's used up, AgentCore keeps serving requests and bills them at the standard rate. The pricing page doesn't mention the allowance yet, so we'd set up the zero-spend budget the docs suggest before relying on it.

By default each account gets 10 requests per second per region, and AWS Support can raise that limit. In a worked example on DEV Community, developer Sascha estimated that 400 questions a day at two searches each comes to about US$168 per 30 days. The same post caps each question at three searches so a looping agent can't run up costs.

Limits to plan around

The acceptable use terms require you to keep and display the citations and links for any result you show users. They also bar bulk extraction or storage of results and building a competing index. If you planned to cache results heavily, design with those terms in mind.

AWS's blog notes that CloudTrail doesn't log InvokeWebSearch unless you add an event selector for AWS::BedrockAgentCore::WebSearch. Even with logging on, the records don't include the search text or filters. The DEV Community post points out that queries still leave your application, so keep personal information out of them. The same post also warns that a date-only publishedBefore value means midnight UTC at the start of that day. To include the whole day, use its last second, as the reference recommends, and convert New Zealand dates to UTC first.

Coverage quality is still an open question. The DEV Community author doubted how well the index handles German-language and niche content. We haven't found any published test of how well it covers New Zealand sources or te reo Māori pages. The docs also don't say whether a date filter drops results that have no known publication time.

What builders could do with it

The domain filter suits grounded question answering over a fixed set of trusted sites. Because a root domain matches its subdomains, a single govt.nz entry should cover every agency site under that domain. A compliance or policy assistant could then cite only government pages. Pairing publishedAfter with an include list of vendor domains gives you a cheap release monitor that runs on a schedule. For teams already running Bedrock models in Sydney, the API can sit behind a tool definition in your own Converse loop, with the domain list kept in code where the model can't see or change it.

If you build something with the API, we'd like to see it on the ASAI projects page.

Sources

Build with people working on the same things

Join the Auckland Society for AI: NZ$599 a year with a ForgeBox and the members' projects area. Our Auckland meetups are open to everyone.

Become a member