Ask or Assume? Uncertainty-Aware Clarification-Seeking in Coding Agents Boosts Resolve Rate 13%

arXiv·high signal

Evaluates clarification-seeking abilities of LLM coding agents on an underspecified variant of SWE-bench Verified. Proposes a multi-agent scaffold that decouples underspecification detection from code execution — one agent identifies ambiguity and asks clarifying questions, another executes. OpenHands + Claude Sonnet 4.5 achieves 69.4% task resolve rate vs 61.2% for standard single-agent setup, an 8.2 percentage point improvement. Directly challenges the 'autonomous execution' paradigm — agents that know when to ask outperform agents that always guess.