A late requirement destroys about twice as much agent-written code as an ordinary edit, and warning the agent doesn't help.

Mining 3,553 eligible SWE-chat sessions for requirements arriving after the agent has already implemented something, linked to deletion or replacement of prior agent-authored lines. Arrival is followed by roughly 2x the invalidation of matched non-requirement edits, with no decline over the session and no association with operation type. The controlled experiment found delayed disclosure just relocates implementation to after the reveal, while advance warning produced no detected effect. So "tell the agent up front that requirements may change" is advice with no measured support. arXiv 2609.03028