Google Cloud Open-Sources 8-Component Harness², Lifting GLM-5 by 11.5 Points
Updated
Updated · MarkTechPost · Oct 18
Google Cloud Open-Sources 8-Component Harness², Lifting GLM-5 by 11.5 Points
3 articles · Updated · MarkTechPost · Oct 18
Summary
Harness² is a plug-and-play framework that rewrites prompts, tools, memory and other agent-harness files during deployment while keeping model weights frozen; Google released it as version 0.1.0.
One refinement step improved all 4 model-harness setups across 3 benchmarks, with the biggest gain lifting GLM-5 under OpenCode to 75.7 from 64.2 on WorkBuddy-Bench and the smallest raising JobBench to 24.8 from 23.6.
The system edits 8 harness components through a recursive Propose, Probe, Compose loop, then stores successful changes and lessons in an EXPERIENCE.md library to guide future tasks without labels or test cases.
Google said weaker executors benefited most: after 1 step, GLM-5 topped base Gemini 3.5 Flash on WorkBuddy-Bench, while parallel recursion at k=5 pushed Gemini under OpenCode to 80.8.
Harness² runs on Linux with Python 3.12 or 3.13 and Vertex AI-backed model calls, positioning it as a deployment-time alternative to methods that require training splits or fixed harnesses.