SWE-Touch: Benchmarking Coding Agents When Users Touch the Code

SWE-Touch evaluates coding agents in shared workspaces where users inspect or modify task-relevant code during an ongoing task. The framework introduces validated Counter-Edits: plausible user changes that conflict with completing the target task. A separate User Patch Generator creates these edits, which are injected with contextual messages when agents reach relevant code. Across nine coding models on SWE-bench Verified, Counter-Edits reduce average resolve rate by 7.7 percentage points. Similar degradation appears on longer-horizon tasks from SWE-Bench Pro and DeepSWE. Trajectory analysis attributes failures to weak awareness of evolving workspace state, insufficient reconciliation, and inadequate targeted testing after changes.