Agent context becomes an enterprise layer

AWS and Microsoft are showing the same trend from two directions this week: enterprise agents need a managed context layer. The model alone is no longer the decisive factor. What matters is how reliably documents, web knowledge, permissions, tools, and cost controls enter the workflow.

What AWS announced

AWS introduced Amazon Bedrock Managed Knowledge Base, a managed RAG component for agents. According to AWS, the service brings storage, retrieval, embeddings, re-ranking, and model selection into one managed primitive. For developers, that means less custom work around vector databases, parsers, retrievers, and observability.

The integration with Bedrock AgentCore Gateway is equally relevant. Knowledge bases can be added there as a prebuilt target; AWS describes automatic role-based permissions, policy enforcement, observability, and evaluation metrics. In addition, Web Search on Bedrock AgentCore is now generally available — initially in us-east-1 — and returns snippets, source URLs, titles, and publication dates through an MCP-based gateway. AWS lists the price at 7 US dollars per 1,000 search queries.

What Microsoft emphasizes in parallel

Microsoft made Copilot Cowork generally available worldwide. The service is designed to run long-running, multi-step tasks while combining Work IQ, Microsoft 365 permissions, plugins, model choice, and usage-based Copilot Credits. Here too, the focus is not on a single chat interface. It is controlled execution across business context and tools.

Why this matters for enterprises

For CIOs, the architecture question is shifting: agents are no longer isolated prompt front ends. They are becoming runtime environments with retrieval, web access, actions, cost centers, and audit requirements.

Before the next pilot, review three points: Which sources may the agent use? Which actions may it execute? And how will retrieval, web search, and tool calls be measured, limited, and owned?

The key question is no longer: “Which model is better?” It is: “Is our context layer robust enough for production agents?”