An agent that performs well in a demo may behave differently when models, tools, data, or user behavior change in production.
This session demonstrates how Microsoft Foundry tracing, evaluation, Trace Replay, and Agent Optimizer create a continuous AgentOps lifecycle. Using a production support agent, we’ll trace decisions and tool calls, identify failures, convert real interactions into evaluation datasets, and validate an improved agent before deployment.
We’ll cover quality and safety metrics, regression gates, latency and cost monitoring, human review, and auditability. Attendees will leave with a practical approach for continuously measuring, improving, and governing enterprise agents.