Google open-sources Harness2: agents that refine their own harness, up to 11.5-point gains
chemist_slime · reddit · 2026-10-19
Google Research released Harness² (Apache 2.0), a recursive agent harnessing framework that improves the system around a fixed-weight model — prompts, tools, memory, skills, subagents — with the agent itself iterating.
- Task-level recursion: a contrastive Propose–Probe–Compose loop proposes harness edits, probes candidate versions on the same task, and composes a better harness from behavioral contrasts — no ground truth or graders needed.
- Harnessing-level recursion: adopted harnesses and evidence-backed lessons persist in an improvement harness, carrying across tasks.
- Results: across 3 professional benchmarks and 4 model-harness configs, one step already beats base harnesses — GLM-5 and Gemini 3.5 Flash gain up to 11.5 points on WorkBuddy-Bench, 4.5 on JobBench, and 11.8pp on LAB partial pass rate; more recursion budget yields more.
- Supports parallel/sequential recursion, OpenCode and Codex executors.
More from coding & agent
- 300 parallel agents plus a 100ms cheap verifier compress weeks of analyst work into one run — CurieuxExplorer · 2026-10-19
- Building a vast Unreal Engine world from scratch with Opus, says AI game dev is accelerating — sven_ai · 2026-10-19
- DeepSeek V4.1-Flash lands #2 on Terminal-Bench 4.0 at 1/16th the price of Opus 5 — zainhas · 2026-10-19
- Hermes Agent + local models manage all machines; hermes-skins repo hits 607 stars — Teknium · 2026-10-19
- Harness is dead? AI Engineer NYC debates stripping scaffolds vs. multi-day agents — TheTuringPost · 2026-10-19
- Open-source Claude Code plugin routes grunt work to cheap models, saves Opus for judgment — investigatormaker · 2026-10-19