Negotiating Completion in Agentic Software Work: Acceptance Boundaries and Deferred Requirements

Coding agents increasingly participate in software development, but submitting an implementation does not settle whether work is complete. This qualitative process comparison examines 30 diagnostic pull requests from 22 GitHub repositories within a frame of 1782 contributions from 120 repositories. Manual screening found substantive contestation in 516 frame contributions (29.0%). The analysis draws on public discussions, reviews, commits, tests, and linked artifacts to trace how participants address unmet or unproven requirements. Using acceptance boundary as a sensitizing distinction, the analysis connects conditions stated in review to their enactment in evaluation artifacts and subsequent revision. Participants may retain requirements for repair, delimit a smaller acceptable increment, pursue an alternative implementation, or terminate an approach. Tests enact these decisions through repair, authorized exclusion, and restoration. In sqruff, issue closure accompanied partial repair and renewed deferral; a subsequent rollback reversed the accepted changes. Integration, issue closure, and requirement satisfaction therefore represent distinct accomplishments. The process model identifies conditional, overlapping paths connecting local acceptance to continuing work. It explains collaborative completion in agent-associated software development and shows how documented interface constraints and human intervention shape subsequent work.