Getting started with Web Answer on Amazon Bedrock AgentCore

AWS announced the general availability of Web Answer on Amazon Bedrock AgentCore. Web Answer is a managed tool that accepts a natural-language question and returns a synthesized answer in seconds, grounded in current web sources with each claim tied to a citation. The pitch is that agent builders no longer have to assemble a search-retrieve-synthesize-cite pipeline themselves.
That pipeline is familiar to anyone who has built retrieval over the live web: call a search API, pick which results to trust, fetch and clean page content, prompt a model to merge it, then engineer citation extraction so output traces back to a URL. Each stage is maintained code and a failure point, whether hallucinated answers on thin evidence, stale pages, or claims drifting from their sources. Web Answer replaces the sequence with one MCP tool call.
The request flow runs through AgentCore Gateway, a capability of Amazon Bedrock AgentCore. An agent sends its question to the Gateway, which authenticates the caller with AWS Identity and Access Management or a JSON Web Token, then calls Web Answer, a connector AWS owns, using its own service role. Web Answer plans searches, runs several in parallel, reads key pages to check facts, searches again when evidence is insufficient, and finally writes an answer mapping each grounded claim to a source URL. The Gateway returns the answer and citations to the agent as a Model Context Protocol tool result.
The design detail worth noting is freshness: Web Answer runs new searches for every question and its model is given today's date, so answers can reflect current information rather than a fixed index. It also carries responsible AI safeguards, per AWS.
Web Answer sits alongside the existing Web Search tools on AgentCore. The dividing line is who runs the model. Web Search returns ranked URLs, titles, snippets and metadata; you keep your own model and control synthesis, which suits teams wanting full control over processing. Web Answer runs the model, the agentic retrieval loop, synthesis and attribution, and hands back a finished cited answer, which suits questions that are sub-tasks inside a larger agent. AWS expects many customers to use both, for example Web Search for broad research needing raw results and Web Answer for specific factual lookups.
Setup is aimed at being quick. A Gateway administrator adds Web Answer as a target on an existing AgentCore Gateway, and the tool then appears to attached agents over MCP via tools/list and tools/call. Prerequisites are an AWS account with AgentCore Gateway access, a Gateway in us-east-1 using IAM inbound authentication, credentials carrying bedrock-agentcore:InvokeGateway on the Gateway, Amazon Bedrock access to the agent's model, and Python 3.10 or later with boto3, strands-agents, mcp and mcp-proxy-for-aws installed.
The walkthrough has three steps. First, grant the Gateway's service role the bedrock-agentcore:InvokeWebAnswer permission, checked against the ARN arn:aws:bedrock-agentcore:us-east-1:aws:tool/web-answer.v1. That role governs only outbound calls; inbound callers are governed separately by the Gateway's authorizer, and the role does not need bedrock:InvokeModel because the agent's model runs under the agent's own identity. Second, create the target with connectorId set to web-answer and credential provider type GATEWAY_IAM_ROLE, so the Gateway calls the backend with the service role and nobody creates or stores credentials for the tool. The target reports READY when it can be called. Third, invoke it from a Strands agent, though any MCP-compatible client works, including LangChain, LangGraph and CrewAI. When a question needs current information, the agent calls the discovered tool, identified as web-answer-tool___WebAnswer.
Beyond the question, the tool accepts optional per-call arguments: a domain filter to include or exclude sites, a published-date range, and a country for localization. A Gateway administrator can also set a domain filter at the target level that applies to every call. Agents can narrow those bounds, excluding more domains or restricting to some included ones, but cannot add a domain the target blocks. This suggests a governance pattern: platform teams set the outer boundary once, and individual agents can only tighten it.
The response is an MCP tool result with both content and structuredContent. The answer is markdown with inline citation markers, and structuredContent.answer carries the same text alongside a structured list of citations, each with a url and, when available, a domain. structuredContent.answer normally holds a single block, and the result includes an isError flag and a latency figure in its metadata. In practice the structured citation list is the more useful half: a product team can render sources in a UI or validate that claims map to domains, without parsing markdown.
Enterprise controls are inherited from AgentCore. There is no separate API key to provision or rotate; callers authenticate to the Gateway, which then calls Web Answer with its service role. Agents can reach the Gateway through an interface VPC endpoint (com.amazonaws.us-east-1.bedrock-agentcore.gateway), keeping traffic off the public internet, and AWS states that queries and results stay within the customer's AWS environment with zero data egress. Operational metrics publish to CloudWatch from the first call. Rate limits are enforced from launch at 10 transactions per second per account in us-east-1, with increases available through AWS Support. Web Answer searches an Amazon-operated web index and Amazon's knowledge graph rather than sending questions to third-party search engines or external websites.
Target use cases named are regulatory Q&A in financial services, clinical support grounded in current literature, customer service where a cited answer beats ten links, legal and compliance research with auditable citations, and internal copilots. Availability is us-east-1 for now, with more regions coming. Pricing is $10 per 1,000 tool calls; following the walkthrough creates billable resources, and cleanup means deleting the target, then the Gateway, then the service role and policy, after which charges stop.
A few trade-offs follow from the design. Because the synthesis model runs under AWS's control, teams lose the ability to swap in a fine-tuned or domain-specialized model, which is exactly what keeps Web Search relevant. Single-region launch and a default 10 TPS cap will matter for high-volume support agents. And a single price per call is simple to forecast but hides variable token costs that a self-built pipeline would expose.
Why it matters: teams building agents that answer questions about current public information can now ship a cited-answer capability with a policy and a target instead of owning a retrieval pipeline, which shifts effort from plumbing to answer quality and source governance. The per-call price and rate limit make cost modeling straightforward, and the administrator-set domain bounds give platform owners a lever over which sources agents may use. The choice between Web Search and Web Answer becomes an architectural decision about who owns synthesis, so expect teams to mix both within one agent.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
Web Answer on Amazon Bedrock AgentCore returns a synthesized, source-cited answer to a natural-language question in a single tool call, with no search, retrieve, synthesize, and citation pipeline to build. Learn how it works and how to add it to an AgentCore Gateway and call it from an agent.