Harness2: Recursive Self-Improving Agent Harnesses Without Verifiers

Ruiyao Xu, Yanfei Chen, Zhongying CuiZhu, Bhavana Dalvi Mishra, Yifei Ming, Han Yu, Rujun Han, Chen-Yu Lee, Tomas Pfister · Google Cloud AI Research · 2026

arXiv 2610.14224

The harness = eight editable files

Agent = frozen model + harness. Hover a tile. Weights never change.

Hover a file tile.

Two coupled loops

Toggle how the second loop feeds future tasks.

Propose → Probe → Compose

JobBench case study: 43.1 → 77.8. Step through one refinement.

Spending the budget k

Parallel = width (one population update). Sequential = depth (chained contrastive steps).

k = 1

Reported results

All numbers are from the episode's discussion of the paper. Higher is better.

Which margins survive the noise?

Three runs, shuffled order, no significance tests. Dots = means, bars = ±1 SD where reported.

Hover rows for detail. “n/r” = SD not reported in the episode.

Evidence gaps Illustrative

Qualitative reading of the episode's critique, not paper data. Blue = covered, red = untested.

covered

untested