AWS adds domain and date filters to Bedrock AgentCore Web Search

What AWS shipped

AWS has added two new controls to the Web Search tool inside Amazon Bedrock AgentCore: a domain allow-list and a publish-date cutoff. According to a dev.to write-up referencing the AWS Machine Learning Blog, AgentCore is AWS's managed runtime for building and running AI agents, and its Web Search tool lets those agents pull live information from the internet rather than depending only on a model's training data or a fixed knowledge base.

How the filters work

Developers can restrict a search to a defined set of domains and set a recency window. Results that fall outside the allow-list, or that are older than the cutoff, are dropped before they ever reach the agent's reasoning step — the model never sees them, rather than being asked to disregard them after the fact.

That distinction is the core of the change. The dev.to post frames the update as a response to one of the messier problems in deploying agents: a model with open-ended web access has no built-in sense of source quality or freshness, and will cite a stale forum thread with the same apparent confidence as an official document. Prompt-level instructions to prefer certain sites or recent content may or may not be followed. Configuration-level filtering moves that judgment upstream, where the builder controls it deterministically.

Where teams are likely to use it

The article points to three operational scenarios where the change matters most:

• Customer support agents that must cite only company-approved documentation. An agent can be pinned to a company's help center plus a short list of vetted vendor sites.
• Sales and research agents that need current pricing or competitive information. A date cutoff excludes anything published before a chosen point, so the agent is not grounding answers in dated pages.
• Workflows with compliance requirements to demonstrate which sources an AI system was permitted to draw from.

There is also a security angle. Restricting the reachable source pool to fewer, known domains narrows the attack surface for prompt injection delivered through malicious web content, since the agent simply never fetches pages from anywhere else.

Scope and open questions

The feature is specific to AgentCore's Web Search tool. It does not change how other retrieval methods, such as vector databases or internal knowledge bases, behave. Teams already running agents on AgentCore with web search enabled can apply the filters through the tool's configuration.

One detail the dev.to post flags as unresolved: AWS's announcement does not mention any pricing changes tied to the update, so any cost impact should be treated as unconfirmed until tested directly.

Why it matters

Grounding agents in live web data is table stakes for many deployments, but unfiltered search pushes the burden of judging source quality onto the model, which is poorly suited to the job. By making domain and recency constraints part of the tool configuration, AWS converts a soft, prompt-level hope into a hard, enforceable boundary. For regulated or customer-facing deployments — where citing the wrong or outdated source carries real consequences — that is the difference between an agent whose citations are merely plausible and one whose source pool can be pinned down and audited. It also reflects a broader pattern in agent design: reliability increasingly comes from constraining the environment a model operates in, rather than from asking the model, politely, to behave.