Microsoft's Self-HPO co-evolves agent harness and model, lifting Qwen3.5 resolve rate from 57.8% to 64.8%
dair_ai · x · 2026-10-21
Microsoft Research and collaborators introduce Self-HPO, a recursive self-improving framework that alternates harness and policy optimization for coding agents:
- Each round, the current policy analyzes its own failures and proposes new harness candidates, then is further trained with RL under the selected harness; tasks that stay hard become the search set for the next harness round.
- Across three coding benchmarks, Self-HPO raises Qwen3.5-35B-A3B's average resolve rate from 57.8% to 64.8%, with joint gains exceeding the sum of harness-only and policy-only improvements.
- Both optimized components retain gains even when paired with the other's base version, and pass@k reliability improves.
- Using Opus 4.8 as the harness proposer adds further gains, showing frontier harness judgment can be transferred to smaller policies.
Code and models are open-sourced (GitHub: MSR-Orchard/Self-HPO).
Related event: Microsoft's Self-HPO Lets Coding Agents Rewrite Their Own Harness(3 posts)→
More from coding & agent
- Dev uses Claude Opus 5.5 to build an interactive playground on Apple hardware ray tracing — dreamwieber · 2026-10-21
- FireTrident lets 80 coding agents share one repo, clearing Hono's 958 violations in 62s — dinasaur_404 · 2026-10-21
- Mark a PyTorch issue "actionable" and 9,999 Claude instances swarm in instantly — generativist · 2026-10-21
- Open-source Lumina Capital: a multi-agent AI hedge fund engine with 7-persona debate council — OpenComplaint524 · 2026-10-21
- Stanford research shows switching agent harnesses mid-run cuts cost and boosts performance — StanfordAILab · 2026-10-21
- NoHumans.Space opens beta: a social knowledge network where AI agents share API findings — dark-epiphany · 2026-10-21