Bedrock AgentCore adds Web Search API: live web grounding in one signed call

AWS has added a Web Search API to Amazon Bedrock AgentCore. Developers call it as one signed request using the AWS SDKs or CLI. There is no gateway to configure and no infrastructure to provision. AWS positions it for applications, agent loops and data pipelines that need fresh information from the web to ground their responses.

The design is deliberately minimal. Many agent frameworks reach the web through third-party search APIs such as Tavily, Serper, Exa or Brave, each with its own keys, billing and rate limits. Using AWS's signed-request model means search falls under the same IAM permissions, logging and billing as the rest of an AWS deployment. For enterprises, that simplifies procurement and security review, which often slows the adoption of outside APIs. It also helps auditing, since searches appear in the account's existing monitoring.

The launch fits AWS's push to make AgentCore a complete runtime for agents. In the past few weeks it has added AgentCore Runtime V2 with faster cold starts, an Agent Registry that syncs with MCP servers, the UAEF evaluation framework, and Nitro Sandboxes for isolating agent-run code. AWS reports tens of thousands of AgentCore customers. Search grounding is a basic requirement for agents that answer current-events questions or research tasks, and offering it natively removes one reason to use an outside platform.

Open questions remain. The announcement does not specify which search index powers results, how ranking works, or how fresh and complete results are compared with Google or Bing-based services. Pricing and rate limits will decide whether it competes with specialist providers on cost. Teams should also consider prompt-injection risks: any web result can contain instructions aimed at the agent, so search output needs the same handling as other untrusted input. Watch for benchmarks against existing search APIs and for integration into Strands and other AWS agent frameworks.