Gate the agent's edit on evidence it actually collected: ECLoop adds 4.8–11.8 points of Pass@1 while cutting tokens 12.1%

arXiv 2607.28815·high signalXBlueskyLinkedInCopy link

Coding agents routinely edit source or submit a patch before examining enough repository evidence to justify the change. ECLoop is an execution layer that compiles evidence conditions from the issue text and repo structure, tracks which the trajectory has satisfied, and postpones any edit or submission whose preconditions are unmet. On SWE-bench Verified (500 instances, two models and two scaffolds) it lifted Pass@1 by 4.8–11.8 points while reducing average token use up to 12.1% — with no fine-tuning and no scaffold changes, and structured conditions beat natural-language ones in ablation.