A single repair plan helps coding agents resolve more tasks

Coding agents can spend much of a repair task inspecting a repository, reproducing a bug and testing before proposing a source-code edit.

The question is whether advice can use that accumulated evidence and arrive before the proposed change is carried out, while the agent still has its investigation and working files.

PAUSE gives a coding agent one repair plan at its first proposed source-code edit, then lets that same agent continue with its gathered evidence.

In recorded repair runs, agents often investigate a repository before proposing a source-code edit. Could a plan use that evidence before the agent commits to a change?

The study inserts one separate Planner at the first proposed source-code edit, then lets the same Executor continue with its conversation and workspace. In SWE-bench Verified, the evaluator found 65.0% of tasks resolved with PAUSE versus 50.2% without guidance.

The matched live comparison supports a benefit in that setup; the aggregate rate does not mean every task improved.

The live results concern the tested repair tasks and agent setup. PAUSE’s trigger depends on an exposed file-editor request for a source-code path; offline recovery also requires a usable checkpoint and restorable workspace. The results do not establish transfer to substantially different repositories or longer workflows.