Google Open-Sources Harness², an Agent Framework That Improves Its Own Harness
algo_diver · x · 2026-10-18
Google Research open-sourced Harness² (2.7k GitHub stars, Apache-2.0), a plug-and-play framework for recursive agent harnessing. A harness is everything around the model — prompts, tools, memory, skills, orchestration. Harness² lets agents refine it without touching weights, via two levels of recursion: task-level recursion uses a contrastive Propose–Probe–Compose step (no ground truth or grader needed, since it learns from behavioral differences between candidate runs), while harnessing-level recursion persists adopted harnesses and evidence-backed lessons to guide future tasks. Results: a single step lifts GLM-5 and Gemini 3.5 Flash by up to +11.5 points on WorkBuddy-Bench, +4.5 on JobBench, and +11.8pp on LAB's partial pass rate. Supports parallel/sequential recursion, OpenCode and Codex executors.
More from coding & agent
- 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
- Developer builds a standalone Mac app from scratch in 1.5 hours with Cursor — gaganghotra_ · 2026-10-19
- Apify founder Jan Curn: every agent payment protocol exists, real volume is near zero — AI Engineer · 2026-10-19
- How I imported 10,000+ media notes into Obsidian with AI, safely — dSebastien · 2026-10-19