AWS adds web search for LangGraph dependency monitoring agents

Search results designed for controlled agent workflows

The API provides current web results through a single call signed with the AWS credentials an application already uses. AWS says every query is served from a web index it operates, rather than being sent to a third-party search engine. This is intended to keep searches inside the AWS environment, a consideration for teams whose queries may reveal internal work.

Each result includes a page title, URL, publication date, and an extract relevant to the query. Applications can include or exclude domains and restrict results to a publication date range. Agents running on AWS Lambda, Amazon Elastic Container Service, or Amazon Bedrock AgentCore runtime can call the API with their execution role, without storing or rotating a separate search API key.

Web Search on Amazon Bedrock AgentCore is also available through AgentCore Gateway. AWS says a single administrator source policy can then cover every agent and Model Context Protocol client.

LangGraph example monitors packages and container images

AWS’s example builds a daily dependency watch for a security team as a LangGraph graph. Each morning, it checks whether new information affects the open source packages and container base images that the team ships. The graph researches npm, PyPI, and container base images in parallel, while code keeps the date window, source blocklist, and search budget outside the model’s control.

Inside each branch, an Anthropic Claude model in Amazon Bedrock decides what to search for and produces structured findings. The search tool adds the fixed date window and blocklist to every query, and allows up to four searches per branch. A run therefore makes at most 12 searches.

After searching, code checks each finding against the returned sources. The cited URL must have appeared in the branch’s results, the source must name the package, any CVE or advisory ID must appear in the source text, and the publication date must fall inside the watch window. LangGraph then compares surviving findings with its store, so previously reported findings are not reported again, before building a cited table for the security team.

AWS reports faster parallel execution in its test

In AWS’s test, the three branches finished in 36 seconds when run in parallel, compared with 99 seconds when run one after another. The run made 12 searches and returned 80 sources. The evidence checks removed two findings, one because its source did not name the package and another because its CVE ID did not appear in the source. Four findings passed the checks, and two were reported because the other two had already been reported in earlier test runs.

AWS cautions that these checks do not confirm severity, exploitation status, affected versions, or the recommended action. They also do not prove when an issue was first disclosed, because a page dated by the index this week can describe an older incident. The company recommends verifying findings against authoritative vulnerability feeds and having someone review a finding before acting on it.

For LangChain developers, the integration offers current, cited search results as data that code can inspect. AWS says the tool works with LangChain agents, LangGraph, and Deep Agents, giving teams a way to combine model-driven research with rules that the model cannot change.