A late requirement destroys about twice as much agent-written code as an ordinary edit, and warning the agent in advance does not help

arXiv 2609.03028·medium signalXBlueskyLinkedInCopy link

arXiv 2609.03028 mines 3,553 eligible SWE-chat sessions for requirements that arrive after the agent has already implemented something, and where repository state can be replayed, links each arrival to deletion or replacement of prior agent-authored lines. A requirement's arrival is followed by roughly twice as much invalidation as matched non-requirement edits, robust to user-turn and net-deletion checks; the burden shows no decline over the course of a session and no association with operation type. A controlled experiment found that delayed disclosure merely relocates implementation to after the reveal, while advance warning produced no detected effect on overwriting, which undercuts the common advice to tell the agent up front that requirements may change.