SWE-Shift Benchmark Tests Coding Agents Under Mid-Execution Environment Changes

SWE-Shift evaluates how coding agents handle evolving workspaces, such as dependency updates or requirement shifts, during execution. Testing 17 systems reveals significant performance degradation; notably, GPT-6 Astra's success rate dropped from 67.35% to 24.79% when faced with task-relevant updates.