How do you find the root cause of AI agent failures?

Debugging AI agents often feels different from debugging traditional software.

A failure might be caused by the model, a tool call, missing context, state management, external APIs, or a chain of decisions made much earlier in the workflow.

When something goes wrong, how do you isolate the actual root cause?

Do you rely on traces, execution logs, replay systems, observability platforms, or custom debugging tools?

I'd love to hear how your team investigates difficult production failures.