RuHarness Evolves the Rubric Alongside the Harness, Gaining 4.5-5.3 Points Over Baselines
dair_ai · x · 2026-10-30
dair-ai highlights RuHarness, a new self-improving agent harness that fixes a key gap in harness-evolution methods: existing approaches evolve the harness but keep the guiding criteria fixed, so they stop matching the agent's weaknesses after a few rounds.
RuHarness co-evolves the rubric with the harness — each round it mines behavioral criteria from trajectories, re-scores them, and directs harness edits at the most relevant unresolved weakness. Criteria that stop firing go dormant and return if the weakness reappears.
Results: it beats the strongest harness-evolution baseline by 4.5 points on held-out tasks with Qwen3.5-35B-A3B and 5.3 points with GLM-4.7-Flash. Takeaway: if harness evolution plateaus, update the critic too.
More from coding & agent
- Astra inside Codex harness shows 'AGI-ness' other frontier models lack — tszzl · 2026-10-30
- CentOS creator builds Ginux kernel in 11 days: Codex wrote 139,000 lines of Rust — JFPuget · 2026-10-30
- What Is an Agent Harness? The 8 Shared Components Behind Claude Code, Codex, and Antigravity — Saboo_Shubham_ · 2026-10-30
- Reddit debate: which small tasks still need human supervision for AI agents — Ok_Army_9681 · 2026-10-30
- Claude keeps asking devs to physically fold the iPhone while coding for Apple's foldable — airkatakana · 2026-10-30
- johnlindquist Praises herdr's Agent-First Desktop App, Already Modding It Heavily — johnlindquist · 2026-10-30