Google Open-Sources Harness²: Letting AI Agents Rewrite Their Own Scaffolding, Hits GitHub Trending

What happened

The core idea behind Harness²: most self-improvement research relies on labeled data or test cases to judge progress, but real deployments usually have neither. Harness² assumes tasks arrive unannounced, no grader exists, and the model stays exactly as shipped — only the scaffolding changes: prompts, subagents, and scripts routing requests to the right tools.

Its mechanism, called Propose-Probe-Compose, is designed for open-ended tasks with no answer key: the system first proposes two candidate edits to its own harness (one for a general weakness, one specific to the task at hand), then probes both by actually running them under matched conditions against a reference run, and finally keeps only the changes the behavioral contrast supports. No labeled correctness signal is required.

Google's team tested the approach across benchmarks, running models like GLM-5 through harnesses like OpenCode and Codex to measure what a single refinement pass could achieve without touching underlying weights. The underlying paper was published on arXiv on October 9, with code at github.com/google-research/harness2. The repository states it is research code, not an officially supported Google product.

Key facts

Open-source details: released October 9, 2026, Apache 2.0 license, version 0.1.0, led by Tomas Pfister with Ruiyao Xu, Yanfei Chen, Rujun Han and others, affiliated with Google Cloud AI Research (not Google DeepMind).

Traction: over 6,000 new GitHub stars in a single week, trending on GitHub in October. Repository: https://github.com/google-research/harness2.

The 8 rewritable files: system prompt, guardrails, memory, tool guidance, subagents, skills, scripts, plugins.