Opera Verbal-Critic Framework Catches Long-Horizon Coding-Agent Failures Mid-Execution

The Opera paper, posted October 9, treats a coding-agent critic as an intervention whose value is revealed afterwards and tackles three design questions: when to intervene, whether a diagnosis is justified, and how to deliver feedback. It builds on Self-Refine, Reflexion, CRITIC and SWE-PRM to judge unfinished work without derailing runs that would otherwise succeed, with experiments showing fewer redundant explorations and premature completion claims on repository-level tasks.