AWS adds Web Search and continuous learning to Bedrock AgentCore

At Summit NYC, AWS significantly expanded Bedrock AgentCore, its platform for building and running production AI agents. The headline addition is a fully managed Web Search tool that lets agents ground their responses in current, cited web knowledge—without the developer manually wiring up a search API, and critically, with zero data egress from the customer's secured environment. For enterprises wary of leaking proprietary context to third-party search services, the no-egress design is a meaningful differentiator.

Beyond web search, AgentCore gained capabilities connecting agents to organizational knowledge, the open web, and paid knowledge sources, plus production-grade debugging and scalable controls aimed at closing the reliability gaps that keep many agent prototypes from reaching production. AWS framed this as 'broader knowledge and continuous learning,' positioning AgentCore as the runtime where agents not only act but improve over time.

The move reflects a clear market read: the bottleneck for enterprise agents isn't raw model capability but grounding, observability, and governance. By bundling managed web search with knowledge connectors and debugging, AWS is trying to make agents trustworthy enough for regulated, high-stakes workloads—the same customers it courts with its database and security portfolio.

Competitively, the web-search-with-no-egress approach contrasts with bolt-on search integrations and directly challenges Microsoft's and Google's agent platforms. It also pairs naturally with the new Managed Knowledge Base, giving AWS a vertically integrated agent stack from data ingestion to grounded action. The open question is real-world citation quality and latency: managed convenience only wins if the grounded answers are accurate and fast. Watch for benchmarks and how pricing for paid-knowledge connectors shakes out as enterprises pilot these tools.