Coding agents lose 7.7 points when you edit files mid-task

Coding agents lose 7.7 points of resolve rate when a human edits their target files mid-task, per a new benchmark called SWE-Touch. I edit files while Claude Code works in the background on almost every project I ship, so this isn't abstract.

SWE-Touch mines task-critical code regions from past repair trajectories. It builds a plausible Counter-Edit that conflicts with the task. Then it drops that edit into the repo with a contextual message right as the agent reaches that region.

The drop holds across nine models on SWE-bench Verified. It isn't one model's quirk, since the same failure pattern persists on SWE-Bench Pro and DeepSWE, detailed in arXiv:2608.02499.

Trajectory analysis found two failure patterns. Some agents keep the conflicting code exactly as it landed. Others overwrite it and move on without re-checking the rest of the repo.

The paper doesn't test whether forcing a fresh repo diff before every write would close the gap. But the failure pattern points at a process problem. Agents that never re-inspect the repo after a conflicting edit lands are the ones losing points, regardless of which model runs the loop.