Amazon Bedrock AgentCore adds web, paid data and feedback

Amazon Bedrock AgentCore now gives agents three knowledge layers, payments and real-time optimization so they can reach internal documents, the live web and paid content while learning from production traces. The release adds Bedrock Managed Knowledge Base, Web Search on AgentCore, AgentCore payments integration with WAF AI traffic monetization, and built-in optimization tools that surface failure, intent and trajectory insights across hundreds of sessions.

What did Amazon add to AgentCore?

AgentCore now natively connects agents to organizational, web and paid knowledge and includes continuous-improvement tooling and stronger policy enforcement. Bedrock Managed Knowledge Base handles ingestion, vector stores, embeddings and re-ranking. Web Search returns high-value web excerpts and combines public sources with Amazon's proprietary knowledge graph. AgentCore payments and WAF AI traffic monetization create a payments and provider-control channel. On the optimization side, AgentCore provides failure, intent and trajectory insights in preview, plus recommendations, batch evaluation and A/B testing to validate fixes.

How do the three knowledge layers work?

The Managed Knowledge Base gives agents native access to internal files by managing vector stores, embeddings and re-ranking so teams do not have to build ingestion pipelines. The platform connects unstructured sources such as SharePoint, Google Drive, Confluence, S3 and internal wikis and uses an agentic retriever that plans queries, connects related concepts across documents and re-ranks results to produce broader, more complete answers than basic retrieval.

Web Search on AgentCore supplies live web information within the customer’s secured AWS environment, using the same Amazon search infrastructure that powers Alexa+, Amazon Quick Suite and Kiro. It is optimized for agentic retrieval and combines public web excerpts with Amazon’s proprietary knowledge graph to add structured entity data and real-time facts such as stock prices and sports scores.

The paid knowledge layer has two parts. AgentCore payments, announced in preview last month, lets agents discover and pay for premium services inside their execution loop. WAF AI traffic monetization, now generally available, lets content owners block, allow or require payment for agent access and recognizes agents verified on AgentCore, creating a trusted channel between agents and providers.

How does monitoring, optimization and control work?

AgentCore turns production traces into a continuous improvement loop: discover recurring failures, generate fixes, validate them and prove they work before rollout. In preview, AgentCore provides rich failure, intent and trajectory insights across hundreds of sessions, surfacing silent behavioral failures, explaining root causes and ranking problems by impact. Investigations return results in minutes and teams can enable daily or weekly monitoring reports. Recommendations and A/B testing are generally available today; recommendations suggest specific prompt and tool-description changes grounded in traces, batch evaluation runs tests against defined datasets to catch regressions, and A/B testing splits live traffic to compare agent versions under production conditions.

Amazon also extended policy controls with Bedrock Guardrails integration, generally available, which evaluates every agent action at the gateway for prompt injection attempts, harmful content and sensitive data exposure. Those checks run outside the agent’s context so the agent cannot see or reason around them.

"Agent improvement is now a continuous loop grounded in data, not trial and error," said Kazumi Matsuda, Senior Manager, AI Promotion Department, FUJISOFT, summarizing how the optimization stack changed their workflow.

Why it matters

Companies building agents rarely lack model capability; they lack reliable access to the right knowledge and a way to spot silent failures in production. AgentCore stitches those pieces together: internal and external grounding, a mechanism for paid data, deterministic gateway checks, and trace-based optimization tools. That combination aims to reduce months of engineering to connect sources and to move fixes from guesswork to validated changes.

What to watch

Monitor adoption signals: whether customers move internal knowledge off bespoke pipelines onto Bedrock Managed Knowledge Base, and whether providers opt into WAF AI traffic monetization now that payments and provider controls interoperate. The next concrete milestones will be broader GA announcements and adoption case studies showing measurable reductions in silent failure rates.