AWS: Bedrock AgentCore Gets Web Search, Payments, and A/B Testing for Agents

AWS has announced a major expansion of Amazon Bedrock AgentCore, the platform for building AI agents, adding broader knowledge and continuous learning capabilities.

Three Knowledge Layers for Agents

AgentCore now offers three knowledge sources. A managed knowledge base connects SharePoint, Google Drive, Confluence, and S3 with an “agentic retriever” that fetches knowledge incrementally. Web search runs on Amazon infrastructure, while AgentCore Payments (preview) enables agents to purchase access to licensed content. This allows an agent to combine internal documents, the open web, and paid sources in a single flow.

Continuous Learning and Control

For production reliability, AWS adds three control features. Monitoring (preview) detects “silent behavioral errors” — cases where an agent fails without an obvious error. Optimization (GA) recommends prompt and tool improvements and introduces A/B testing in production. Guardrails (GA) evaluates every agent action for prompt injection, harmful content, and data exposure, unlike the earlier approach that only checked input and output.

What AWS Is Changing for Content Publishers

Alongside agent features, AWS WAF introduces AI traffic monetization (GA) — a mechanism that lets content providers charge agents for access. This fits into the broader trend of turning crawling from a free interaction into a billable one. Security partners Check Point, Zscaler, Rubrik, Netskope, and SentinelOne were also announced.