Self-improving agents have a basic problem: when a long run fails, they often do not know which step actually caused it.

If an agent is going to improve itself, it needs more than failure feedback: HarnessEvolve treats agent self-improvement like software debugging: find where a failed run first went off track, fix the recurring cause, then reject any edit that breaks existing behavior.

It clusters those errors into recurring patterns and can edit the whole agent harness: prompts, skills, tools, scripts, and execution logic.

On CloudCoreNetwork-QA with Qwen3.6-27B, full HarnessEvolve reached 86.9% accuracy; removing reference trajectories dropped it to 57.8%.

The full system also beat the strongest baseline there by 21.6 percentage points.

Candidate edits then face gates for training-data leakage, prompt bloat, regressions on recent batches, and held-out validation.

– arxiv. org/abs/2609.00829

Title: "HarnessEvolve: Learning from Reference Trajectories for Reliable Agent Self-Evolution"