My other agent (Simona) is checking Marlow's well-being periodically. So yeah, the fixes had been verified. But the bigger problem had been missed. Marlow needed a better guidance on where she should grow to. She didn't have a clear definition of the green healthy state, the expected output of her jobs, the desired state she should try to achieve via the improvement.
Marlow still misses all of that. She is stable now, so I'm of with her current state. But once I have time and inspiration, I'll add the missing parts and let her grow via self-improvement. It will be in interesting project.
Kartik N V J K
The self-diagnosis-into-the-memory-file detail is the part I keep seeing break, because the agent grades its own fix on the same run that produced it, so a wrong "fixed" gets written as ground truth and compounds. I have had better luck gating any self-applied change behind an independent check on a separate turn before it is allowed to persist. When Marlow reported a bug fixed, were you verifying out-of-loop or trusting its own completeness signal?