Google's HARNESS² evolves agent harnesses without verifiers, up to +11.5 points on WorkBuddy-Bench
dair_ai · x · 2026-10-15
dair-ai highlights Google's paper HARNESS²: Recursive Agent Harnessing for an Open World (Google Cloud AI Research + Northwestern).
- Problem: an LLM agent's effectiveness depends on its harness (prompts, memory, skills, tools), but open-world tasks lack ground-truth verifiers to guide online harness edits.
- Two-level recursion:
- Task-level: a contrastive Propose–Probe–Compose step generates harness edit candidates, tests them against each other, and keeps the best — no verifier needed; scales via parallel/sequential recursion.
- Harnessing-level: lessons from those edits update a continual improvement harness carried to future tasks.
- Results: improved GLM-5 and Gemini 3.5 Flash in every configuration across WorkBuddy-Bench, JobBench and LAB — up to +11.5 points on WorkBuddy-Bench, +4.5 on JobBench, +18.8 points on LAB; running 5 parallel candidate edits adds another 2.9–3.9 points, beating the ReasoningBank memory baseline.
- Discusses recursive self-improvement (RSI) for open-world tasks; code open-sourced at github.com/google-research/harness2.
Related event: Google's Harness²: Recursive Self-Evolving Agent Framework(2 posts)→
More from coding & agent
- Teknium ships official Hermes Agent plugin for one-install reverse engineering of any app — alexcovo_eth · 2026-10-16
- Claude eng lead Felix Rieseberg publishes tutorial on being more ambitious with AI — danshipper · 2026-10-16
- Qwen Code Desktop v0.25.2 ships managed-agent fixes and MCP config repair — github-actions[bot] · 2026-10-16
- NVIDIA makes 7 AI agent courses free, from agent basics to securing agents — ifioknkem · 2026-10-16
- Claude runs 9-hour task, then stalls waiting for Drive permission — juliey4 · 2026-10-16
- vegaml ships a full open-source toolbelt: cascaded inference, calibration, MCP server — kalyan_kpl · 2026-10-16