Web Answer on Amazon Bedrock AgentCore: When a Managed Cited-Answer Tool Beats Your Own Search Pipeline
Prompt & Circuit Editorial
November 9, 2026 · 12 Min Read

Why this matters
Agents that need current information usually carry a retrieval pipeline of their own: call a search API, pick results, clean page content, prompt a model to synthesize, and extract citations [3]. AWS announced general availability of Web Answer on Amazon Bedrock AgentCore on November 5, 2026 [1]. It is a managed tool that returns a synthesized, grounded, cited answer to a natural-language question in a single call [1]. For architects, the question is not whether it works but where it belongs: which questions go to a managed answer tool, which stay with your own retrieval, and what governance you inherit.
What Web Answer is
Web Answer is a managed connector that you add as a target to an AgentCore Gateway [2]. Agents discover it with a standard tools/list call and invoke it like any other Model Context Protocol (MCP) tool [2]. Behind the call, it runs an agentic search: it searches the web, reads relevant pages to ground its answer, and includes citations [2]. It plans its own searches, runs several in parallel, and decides when it has gathered enough evidence [1]. Your agent's model does not need to process raw search results [2].
The tool works with Strands Agents, LangChain, LangGraph, CrewAI, or any MCP-compatible client [2]. When the gateway uses semantic search, tools/list also returns a gateway search tool that can come first in the list, so code should match the tool name ending in ___WebAnswer rather than rely on position [2].
Web Answer or Web Search
The two tools split on who runs the model [3]. Web Search returns ranked results with titles, URLs, snippets and publication dates, and your own model reasons over them [2]. Web Answer runs the retrieval loop, synthesis and attribution for you and returns a finished, cited answer [3].
AWS positions Web Answer for a factual sub-question inside a larger task [2]. Web Search suits cases where you want full control over how results are processed [3]. Many customers will use both: Web Search for broad research and Web Answer for specific factual questions [3].
| Decision factor | Web Search | Web Answer |
|---|---|---|
| What the agent receives | Ranked results with snippets and dates [2] | A synthesized answer with citations [2] |
| Who synthesizes | Your own model [3] | The managed tool [3] |
| Published price | $7 per 1,000 queries [5] | $10 per 1,000 tool calls [5] |
The units differ: Web Search is billed per query submitted, while Web Answer is billed per tool call [5]. Because Web Answer plans and runs several searches inside one call [1], compare cost per finished answer rather than the two list prices alone.
Identity and access
Setup has two parts. First, the Gateway service role needs the bedrock-agentcore:InvokeWebAnswer permission, which the Gateway checks on every request against an AWS-owned ARN [3]. Second, you add a target with connectorId: "web-answer" [2].
Outbound authentication is automatic for connectors: with GATEWAY_IAM_ROLE, the Gateway calls the backend with the service role, so you create and store no credentials for the tool [3]. The role controls only the Gateway's outbound calls, and who can call the Gateway is set by its inbound authorizer [3]. The Gateway can authenticate callers with AWS IAM or a JSON Web Token [3]. The role does not need bedrock:InvokeModel, because the agent's own model runs under the agent's identity [3].
To give an AgentCore harness access, attach the gateway as an agentcore_gateway tool and grant the harness execution role bedrock-agentcore:InvokeGateway on the gateway ARN [2].
Governing what the agent can read
Domain filtering is the main control. In connector version 1.1.0 and later there are two layers that compose on every request [2]:
- Target-level include and exclude lists, up to 50 domains per list, which the agent cannot see [2].
- Request-level filters the agent may pass: domain lists, a published-date range and a country [2].
The composition rules matter for design [2]:
- A domain is excluded if it appears on either exclude list, and a root domain also matches its subdomains.
- If both include lists are set, only domains present in both are used; if they share no domains, the request fails with a validation error.
- Request-level filters cannot override the target-level exclude list or expand beyond the target-level include list.
In practice, the platform team sets the outer boundary at the target, and the agent can only narrow it [3]. On version 1.0.0 the schema exposes only query [2], so confirm the connector version before designing around filters.
Privacy and network posture
AWS states that questions are served within AWS infrastructure: Web Answer searches an Amazon-operated web index, writes the answer with a model on Amazon Bedrock, and does not send questions to a third-party search engine [2]. Processing uses geographic cross-Region inference and remains in supported commercial AWS Regions within the geographic boundary [2]. Agents can reach the gateway through an interface VPC endpoint, com.amazonaws.us-east-1.bedrock-agentcore.gateway, so traffic stays off the public internet [3]. Operational metrics publish to Amazon CloudWatch from the first call [3].
Limits and constraints
- Region: the connector is available in US East (N. Virginia),
us-east-1, and AWS says more Regions are coming [2][3]. The announcement also points to the AWS Region table [1]. - Rate limit: 10 transactions per second per account in
us-east-1, with increases requested through AWS Support [3]. - Input:
queryis 1 to 30,000 characters, and each domain list holds 1 to 50 entries [2]. - Errors: when the tool cannot answer,
isErroris true and the text content describes the problem [2]. - Response handling:
structuredContent.answeris a list of blocks, and the documentation says to read every block [2]. - Acceptable use: you must retain and display the source citations and links with any output that uses a Web Answer result, and you may not use it to extract content in bulk or to build a competing index or database [2].
- Reliability of content: the documentation says Web Answer is not a substitute for professional advice and that outputs should be reviewed before use in regulated contexts [2].
The rate limit and single-Region scope are the two constraints most likely to shape a rollout. A multi-Region agent fleet needs to decide whether cross-Region calls to us-east-1 are acceptable, and a high-traffic agent needs to budget against 10 transactions per second.
Decision guide
Choose Web Answer when:
- the question is a discrete factual sub-task and you do not want to own a retrieval pipeline [3];
- the agent's own model should not spend context on raw search results [2];
- you need citations your agent can show to users [2].
Keep Web Search, or your own retrieval, when:
- your model needs to reason over the raw ranked results [2];
- you want full control over processing and synthesis [3].
Architect's checklist
- Confirm your gateway and agents can operate in
us-east-1before choosing Web Answer [2]. - Grant
bedrock-agentcore:InvokeWebAnsweron the AWS-owned ARN only, and keep the gateway's inbound authorizer as the real access control [3]. - Set target-level include and exclude lists for any agent that must stay inside approved sources, and check the connector version supports them [2].
- Plan for the 10 transactions per second per-account limit and request an increase if needed [3].
- Surface the citations to end users, as the acceptable use terms require [2].
- Handle
isError: trueand read every block instructuredContent.answer[2]. - Estimate spend per finished answer using 7 per 1,000 queries plus your own model's synthesis cost [5].
- Add human review before relying on outputs in regulated contexts [2].
Sources
Numbers in square brackets refer to these sources, checked on 2026-11-09.
Responses (0)
Sign in to join the conversation and share your thoughts.

Prompt & Circuit Editorial
Contributing Editor
Articles on AI infrastructure for cloud architects, written from official vendor sources and checked claim by claim before publishing.
Continue Reading

MCPRoute on Application Gateway for Containers: One Governed Front Door for MCP Servers
Application Gateway for Containers can now route and aggregate Model Context Protocol servers running in Kubernetes. Here is what the preview changes, where it stops, and what to check before adopting it.

Keeping Amazon Bedrock Traffic Private with VPC Endpoints
Which Bedrock APIs can be reached through AWS PrivateLink, how private DNS changes your code, and why the default endpoint policy deserves attention.

From "Just a Chatbot" to a Full AI System: The Layers Explained
Modern AI products are not a single model behind a text box. A layer-by-layer guide to the stack: LLMs, retrieval-augmented generation, tools and agents, the Model Context Protocol, and the runtime that keeps it all reliable.