Web Answer

Web Answer on Amazon Bedrock AgentCore takes a natural-language question and returns a synthesized answer with citations to current web sources. It is a managed connector that you add as a target to your AgentCore Gateway. Agents discover it with a standard tools/list call and invoke it like any other Model Context Protocol (MCP) tool.

Web Answer runs an agentic search for each question. It searches the web, reads relevant pages to ground its answer, and includes citations so you can verify claims. Your agent makes one tool call and gets back a finished, cited answer. You don’t need to build or maintain a search, page-fetch, and synthesis pipeline, and your agent’s model doesn’t need to process raw search results.

Web Answer and Web Search Tool complement each other:

• Use Web Search Tool when your agent’s own model should reason over ranked results (titles, URLs, snippets, and publication dates).

• Use Web Answer when your agent needs a finished answer with citations, for example for a factual sub-question inside a larger task.

Key capabilities

Web Answer provides the following key capabilities:

• Built-in agentic web search – Runs several searches in parallel to cover the question, and reads relevant pages to ground its answer, all from one tool call.

• Answers with citations – Returns markdown text with inline citation markers, plus a structured list of the cited source URLs.

• Current sources – Runs new searches for every question.

• Zero infrastructure management – You don’t provision search APIs, model endpoints, or outbound credentials. The Gateway exposes Web Answer as a standard MCP tool.

• Framework agnostic – Works with Strands Agents, LangChain, LangGraph, CrewAI, or any MCP-compatible client.

• Domain, date, and location controls – In connector version 1.1.0 and later, you can set a target-level domain include list and exclude list, and your agent can pass per-request domain filters, a published-date range, and a country.

Private by design

Before you let an agent send questions to the web, you need to know where those questions go. Web Answer keeps that answer simple.

Questions stay in AWS

When your agent calls Web Answer, the question is served entirely within AWS infrastructure. Web Answer searches a web index that Amazon builds and operates, and writes the answer with a model on Amazon Bedrock. It doesn’t send your questions to a third-party search engine. The Gateway authenticates to the AWS-owned connector and routes the request internally, so the data path stays inside AWS end to end.

Processing stays in your geography

Web Answer uses geographic cross-Region inference, so data processing remains in supported commercial AWS Regions within the geographic boundary. For more information, see Availability.

How it works

Web Answer is a built-in connector. The Gateway handles schema management, parameter governance, endpoint resolution, and service authentication.

1. Gateway setup – You create a Gateway and add a Web Answer target with connectorId: "web-answer". The Gateway snapshots the tool schema and provisions the integration.

2. Tool discovery – Your agent calls tools/list on the Gateway endpoint and discovers the <target-name>___WebAnswer tool with its input schema.

   Note

   If your gateway uses semantic search (searchType: "SEMANTIC"), tools/list also returns the gateway’s x_amz_bedrock_agentcore_search tool, and that tool can come first in the list. If your code selects the Web Answer tool itself, match the tool name that ends with ___WebAnswer. Don’t rely on its position in the list.

3. Question – Your agent calls tools/call with a natural-language question and, in connector version 1.1.0 and later, optional request-level filters and a country. The Gateway authenticates to the backend with its service role and passes on the target’s settings.

4. Agentic search – Web Answer searches the web, reads the most relevant pages, and writes an answer from the evidence that it collected.

5. Cited answer – Your agent gets the answer as markdown text with inline citations, and as structured content with a list of the cited URLs.

Use cases

You can use Web Answer for the following use cases:

• Factual sub-questions – Answer a specific question inside a larger agent workflow without adding a retrieval pipeline to the agent.

• Regulatory and policy Q&A – Answer questions about regulations, standards, or policy changes, with citations so you can verify claims.

• Customer support – Answer product and policy questions from current public information, with sources that your support agent can show to the customer.

• Research assistants – Give knowledge workers current, cited answers that they can audit.

Web Answer is not a substitute for professional advice. Review outputs before you rely on them in regulated contexts.

Set up Web Answer

For instructions on creating a Gateway Target with the Web Answer connector configuration, including setup examples that use the Python SDK and the AWS CLI, see Set up Web Answer in the target configuration guide.

Use with a harness

To give an AgentCore harness access to Web Answer, attach the gateway that you created in Set up Web Answer to the harness as an agentcore_gateway tool. Grant the harness execution role the bedrock-agentcore:InvokeGateway permission on the gateway ARN. For the end-to-end walkthrough, see Tools in the harness documentation.

Configure domain filtering

In connector version 1.1.0 and later, Web Answer supports 2 layers of filtering that compose on every request:

• Target-level domain include list and exclude list. When you set up the target, you configure parameterValues.domainFilter.include and parameterValues.domainFilter.exclude, up to 50 domains per list. Your agent can’t see these lists, and Web Answer applies them to every request that flows through the target.

• Request-level filters. Your agent can pass an optional filters object with each tools/call request:

  • domainFilter – include and exclude lists, up to 50 domains per list.

  • publishedDateFilter – from and to bounds, inclusive, ISO-8601 date or date-time, interpreted in UTC. A date-only to covers the whole day.

Request-level filters compose with the target-level lists:

• A domain is excluded if it appears on either the target-level exclude list or the request-level exclude list. A root domain also matches its subdomains.

• If both the target-level and request-level include lists are set, only domains present in both are used. If the two lists share no domains, the request fails with a validation error.

• If publishedDateFilter.from is later than publishedDateFilter.to, 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.

For configuration examples, see Configure domain filtering in the target configuration guide.

Configure the Gateway Service Role

The Gateway needs a service role that allows the AgentCore service to perform actions on your behalf. For the required IAM permissions and policy configuration, see Configure the Gateway Service Role in the target configuration guide.

Input schema

Web Answer accepts the following input schema when invoked through tools/call. The filters and userLocation fields are available in connector version 1.1.0 and later. On version 1.0.0, the schema exposes only query.

{
  "inputSchema": {
    "type": "object",
    "properties": {
      "query": {
        "description": "The question or search query to answer.",
        "type": "string"
      },
      "filters": {
        "description": "Per-call runtime filters set by the caller for this specific request. Contains an optional runtime domainFilter (merged with the admin-level domainFilter) and an optional publishedDateFilter.",
        "type": "object",
        "properties": {
          "domainFilter": {
            "description": "Per-call domain filter. Include and exclude lists each hold up to 50 domains. Merged with the admin-level domainFilter; the merged exclude list holds up to 100 domains.",
            "type": "object",
            "properties": {
              "include": {
                "description": "Domains to include in search results. When set, results are restricted to these domains (subject to merge rules with the peer filter).",
                "type": "array",
                "items": {"type": "string"}
              },
              "exclude": {
                "description": "Domains to exclude from search results. Root domain matches subdomains (for example, example.com matches blog.example.com).",
                "type": "array",
                "items": {"type": "string"}
              }
            }
          },
          "publishedDateFilter": {
            "description": "Per-call published-date filter. Both bounds inclusive, UTC, ISO-8601. Applies to web results only.",
            "type": "object",
            "properties": {
              "from": {
                "description": "Earliest publication date to include (inclusive).",
                "type": "string"
              },
              "to": {
                "description": "Latest publication date to include (inclusive).",
                "type": "string"
              }
            }
          }
        }
      },
      "userLocation": {
        "description": "Optional. Per-call caller location that may allow web results to be refined for that location. Only country is supported.",
        "type": "object",
        "properties": {
          "country": {
            "description": "Optional. ISO 3166-1 alpha-2 country code, for example \"DE\". Case-insensitive.",
            "type": "string"
          }
        }
      }
    },
    "required": ["query"]
  }
}

Field	Type	Required	Description
query	string	Yes	The question or search query to answer. 1 to 30,000 characters.
filters	object	No	Optional request-level filters. Available in connector version 1.1.0 and later.
filters.domainFilter.include	array of strings	No	Domains to include. 1 to 50 entries.
filters.domainFilter.exclude	array of strings	No	Domains to exclude. A root domain also matches its subdomains. 1 to 50 entries.
filters.publishedDateFilter.from	string	No	Earliest publication date, inclusive. ISO-8601 date (YYYY-MM-DD) or date-time (YYYY-MM-DDTHH:mm:ss with optional Z or +HH:mm offset).
filters.publishedDateFilter.to	string	No	Latest publication date, inclusive. ISO-8601 date (YYYY-MM-DD) or date-time (YYYY-MM-DDTHH:mm:ss with optional Z or +HH:mm offset).
userLocation	object	No	Optional caller location to refine web results. Available in connector version 1.1.0 and later.
userLocation.country	string	No	ISO 3166-1 alpha-2 country code, for example DE. Exactly 2 letters, case-insensitive.

Response format

Web Answer returns an MCP tool result with content, structuredContent, isError, and _meta fields:

{
  "isError": false,
  "content": [
    {
      "type": "text",
      "text": "Under Basel III, banks must hold Common Equity Tier 1 capital of at least 4.5% of risk-weighted assets [[1](https://example.com/basel-iii)]..."
    }
  ],
  "structuredContent": {
    "answer": [
      {
        "text": "Under Basel III, banks must hold Common Equity Tier 1 capital of at least 4.5% of risk-weighted assets [[1](https://example.com/basel-iii)]...",
        "citations": [
          {"domain": "example.com", "url": "https://example.com/basel-iii"}
        ]
      }
    ]
  },
  "_meta": {
    "latencyMs": 5210
  }
}

• content[0].text is markdown text, not JSON. Inline citation markers use the form [[N](url)]. N starts at 1, and a marker can repeat when several sentences cite the same source.

• structuredContent.answer is a list of answer blocks. Each block has text and an optional citations list. The marker in a block’s text refers to citations[N-1] of that block. Read every block in the list.

• Each citation has a url and, when available, a domain.

• _meta.latencyMs reports the server-side latency.

• When the tool can’t answer, isError is true and content[0].text describes the problem.

Field	Type	Required	Description
content[].text	string	Yes	The answer as markdown text with inline citation markers.
structuredContent.answer[].text	string	Yes	The text of an answer block.
structuredContent.answer[].citations[].url	string	Yes	URL of a cited source.
structuredContent.answer[].citations[].domain	string	No	Domain of a cited source.
isError	boolean	Yes	true when the result describes an error instead of an answer.

Availability

The Web Answer connector is available in the following AWS Region:

• US East (N. Virginia) us-east-1

Web Answer uses geographic cross-Region inference. Data processing remains in supported commercial AWS Regions within specified geographic boundaries, such as the US, EU, and APAC. For more information, see Geographic cross-Region inference in the Amazon Bedrock User Guide.

Because Amazon operates the full Web Answer stack, improvements to search, reading, and answer quality reach your agents through the same managed connector. You don’t need to upgrade or migrate anything.

Acceptable use

If you use Web Answer, you are responsible for your use, and any use by your end users, of Web Answer results ("Web Answer Results"). You must retain and display the source citations and links provided with each Web Answer Result in any output you surface to your end users that uses the Web Answer Result. You may not use Web Answer to (a) extract, store, or reproduce content from Web Answer Results in bulk, or (b) build or populate a competing index or database.