Code with Cosmos: Operating AI Agents in Production
Microsoft positions Azure Cosmos DB as unified data foundation for observing, evaluating, and optimizing production AI agents.
“Before you can fix anything, you have to understand what your agent is actually doing.”
Microsoft's Azure Cosmos DB team presented a three-stage lifecycle — observe, evaluate, optimize — for running AI agents in production, with Cosmos DB serving as the single persistent store for operational data, vector embeddings, and agent memory. The session highlighted LangSmith integration, OpenTelemetry-based tracing via Microsoft Foundry, and LLM-as-a-judge patterns for automated regression testing in CI/CD pipelines. The engineering insight is practical but primarily serves as product marketing for Cosmos DB rather than a novel industry-wide signal.