From GitHub · · 1m
GitHub - google-research/harness2
Harness² improves AI agents with recursive task-level and harness-level recursion.
In brief
Harness² improves AI agents with recursive task-level and harness-level recursion. Harness² boosts AI agent performance by 11.5 points on WorkBuddy-Bench through recursive task-level and harness-level optimization. Originally reported by GitHub.
What is Harness²
Harness² is a plug-and-play framework for recursive agent harnessing in an open world.
Task-level recursion
Task-level recursion refines execution harness through propose probe compose steps without ground truth feedback.
Harnessing-level recursion
Harnessing-level recursion updates persistent improvement harness with adopted harnesses and evidence supported lessons.
Performance gains
Harness² improves GLM-5 and Gemini 3.5 Flash by up to 11.5 points on WorkBuddy-Bench and 11.8 percentage points on LAB.
Components
Harness² exposes eight editable components: system prompt, guardrails, memory, tool guidance, subagents, skills, scripts, and plugins.
Installation
Install with Linux, Python 3.12, and uv to run local demos without model credentials.
Benchmarking
Run benchmarks like JobBench, LAB, and WorkBuddy-Bench with harness2 commands for model-backed evaluations.
The takeaway
Harness² boosts AI agent performance by 11.5 points on WorkBuddy-Bench through recursive task-level and harness-level optimization.
