How to evaluate AI agents with tracing, replay, and failure detection so you can catch regressions, safety issues, and broken tool calls before production does.