Switch-SWE: Do Coding Agents Hold Focus When Switching Between Unfinished Tasks?
Across seven coding agents, switching between unfinished tasks lowered completion rates by an average of 11.5 percentage points.
Switch-SWE tests 200 focus tasks from 19 Python repositories in matched uninterrupted and task-switching sessions, scoring only the final focus patch. Completion gaps ranged from 6.0 to 15.0 percentage points across agents. Trajectory analysis found recurring confusion between tasks’ requirements and loss of task state. A context-management plug-in that selects relevant history and keeps concise task summaries improved completion on OpenCode by 5.5 to 15.5 percentage points across evaluated models.
HF Daily Papers' note
Switch-SWE tests 200 focus tasks from 19 Python repositories in matched uninterrupted and task-switching sessions, scoring only the final focus patch. Completion gaps ranged from 6.0 to 15.0 percentage points across agents. Trajectory analysis found recurring confusion between tasks’ requirements and loss of task state. A context-management plug-in that selects relevant history and keeps concise task summaries improved completion on OpenCode by 5.5 to 15.5 percentage points across evaluated models.
HF Daily Papers' note
score 4