Great question. I think logs for AI agents need to be much more structured than normal logs. Trace IDs, clear error context, input values, expected result, actual result, and safe steps to reproduce would help a lot. Without that, the agent may just make guesses from incomplete information.