Google open-sources Harness² to boost AI agents without retraining

Google has released an open-source framework called Harness² that improves the performance of AI agents without retraining the underlying model, according to aitimes.com. The company published the framework on GitHub.

Instead of adjusting model weights, Harness² refines the agent's execution environment: system prompts, tool-use guidance, memory, procedures and other environment files. It uses a Propose-Probe-Compose contrastive method that compares two proposed modifications on the same task without a separate answer verifier. At the harness level, the framework stores lessons from earlier tasks in a library called LESSONS and proposes fixes for new tasks starting from the base harness.

Google evaluated Harness² on WorkBuddy-Bench and JobBench, and on Harby's LAB via pass rates, using GLM-5 and Gemini 3.5 Flash across several harness configurations. A single improvement round raised scores on all three evaluations for all four model-harness configurations tested.

With GLM-5 paired with OpenCode, WorkBuddy-Bench rose from 64.2 to 75.7 points, JobBench gained up to 4.5 points and LAB pass rates rose by up to 11.8 percentage points. Gains were uneven: with GLM-5 and Codex, JobBench moved only from 23.6 to 24.8 points.

Older approaches to improving AI performance often involved retraining models or relying on evaluation data with known correct answers, the kind of setup Harness² is designed to avoid.