Amazon Bedrock AgentCore launches managed Web Answer tool for cited responses

Amazon Web Services has announced the general availability of Web Answer on Amazon Bedrock AgentCore. This new managed tool allows AI agent builders to retrieve synthesized, source-cited answers to natural language questions through a single tool call, removing the need to construct complex retrieval-augmented generation pipelines manually.

What happened

Building agents that answer questions using current web data traditionally requires assembling a fragile chain of components. Developers must call search APIs, filter results for trustworthiness, retrieve and clean page content, prompt a model to synthesize the information, and engineer logic to extract citations so outputs trace back to their sources. Each step introduces maintenance overhead and potential failure points, such as hallucinations when evidence is weak or broken attribution when claims drift from their sources.

Web Answer collapses this entire workflow into one request. When an agent sends a question to the AgentCore Gateway, the system authenticates the call and forwards it to the Web Answer connector. This managed service plans its own searches, runs them in parallel, reads key pages to verify facts, and iterates if evidence is insufficient. It then returns a grounded answer with inline citations, all within seconds.

The service is designed to handle the agentic retrieval loop internally. Unlike raw search tools that return lists of links, Web Answer performs the synthesis and attribution itself. This shift moves the burden of pipeline engineering from the developer to the AWS infrastructure, aiming to reduce latency and improve reliability for knowledge-intensive workflows.

How it works

Web Answer operates as a managed connector owned by AWS, accessed through the AgentCore Gateway. The gateway handles authentication via AWS Identity and Access Management (IAM) or JSON Web Tokens, while the connector uses its own service role to execute searches. The underlying mechanism involves an agentic search process where the system plans multiple queries, retrieves content from an Amazon-operated web index, and validates facts before generating a response.

The output follows the Model Context Protocol (MCP) standard, returning both unstructured markdown text with inline citation markers and a structured content object. This structured data includes the answer text alongside a list of citations, each containing the source URL and domain. Because the service runs new searches for every question and receives the current date, it prioritizes up-to-date information over stale indexed data.

Developers can configure domain filters at the target level to include or exclude specific sites, ensuring agents only draw from trusted sources. These filters can be further narrowed by the agent at runtime but cannot be widened beyond the administrator’s initial bounds. This architecture ensures that queries and results remain within the AWS environment, with no data egress to third-party search engines.

Key details

• Single tool call: Replaces multi-step search, retrieve, and synthesize pipelines with one API request.
• Agentic retrieval: The system plans parallel searches, reads pages for verification, and retries if evidence is insufficient.
• Cited output: Returns grounded answers with inline markers and a structured list of source URLs and domains.
• Enterprise controls: Supports VPC endpoints via AWS PrivateLink, IAM authentication, and admin-defined domain filters.
• Pricing: Costs $10 per 1,000 tool calls.
• Availability: Currently available in the US East (N. Virginia) region, with a rate limit of 10 transactions per second per account.

Why it matters

For software engineers building AI agents, Web Answer reduces the operational complexity of maintaining reliable retrieval systems. Custom RAG pipelines often require significant effort to handle edge cases like conflicting sources or outdated content. By offloading the synthesis and citation logic to a managed service, teams can focus on higher-level agent orchestration rather than low-level search engineering.

The distinction between Web Answer and existing Web Search tools is critical for architectural decisions. Web Search returns raw ranked URLs and snippets, leaving synthesis to the developer’s model. This is ideal for broad research tasks where full control over processing is needed. In contrast, Web Answer provides a finished answer, making it suitable for sub-tasks within larger agents where speed and verified facts are prioritized over raw data exploration.

This capability is particularly relevant for regulated industries like financial services, healthcare, and legal compliance. In these sectors, every claim must be auditable and traceable to an authoritative source. The built-in responsible AI safeguards and strict citation mapping help meet these requirements without requiring custom guardrail implementations.

What you can do

• Evaluate use cases: Identify agent sub-tasks that require quick, factual answers with citations rather than broad research results.
• Set up permissions: Grant your AgentCore Gateway service role the bedrock-agentcore:InvokeWebAnswer permission using an inline IAM policy.
• Configure targets: Add Web Answer as a connector target in your AgentCore Gateway, applying domain filters to restrict sources to trusted domains.
• Integrate via MCP: Use MCP-compatible clients like Strands Agents, LangChain, or LangGraph to discover and call the tool through the gateway.
• Test citation accuracy: Run queries against known factual topics to verify that the returned citations correctly map to the claims in the answer.
• Monitor costs: Track usage via Amazon CloudWatch metrics, keeping in mind the $10 per 1,000 calls pricing model.