Debugging an AI agent that runs for dozens of steps: reading files, calling APIs, writing code, and revising its own output, is not like ...