Termination Lag: Coding Agents Solve Early but Stop Late
Summary
- Researchers characterize termination lag, defined as the steps between an agent's first correct patch and the end of its run.
- Analysis across five models, two frameworks, and three SWE-bench datasets shows 38% to 82% of successful runs find a correct solution on their initial edit.
- Post-solution steps account for a median of 35% to 60% of successful runs, largely spent on verification and repository exploration.
- In 1% to 8% of cases, continued post-solution activity leads agents to undo their correct patch and end with an incorrect solution.
Why it matters
Understanding termination lag highlights inefficiencies and regression risks in autonomous coding agents, showing that continued execution after finding a solution can waste compute and occasionally break working code.