Google researchers open-source Harness², which lets an agent rewrite its own prompts, tools and memory without touching model weights: one step lifts GLM-5 and Gemini 3.5 Flash by up to 11.5 points on WorkBuddy-Bench

Why it matters

【Key points】

Google Cloud AI Research and Northwestern University posted "Harness^2: Recursive Agent Harnessing for an Open World" on arXiv (2610.14224, October 9) and released the code under Apache 2.0 on GitHub. A "harness" is everything around the model: system prompt, tool descriptions, memory, skills, subagents and scripts. Harness² keeps model weights fixed and lets an agent rewrite that layer on unseen tasks.

Task level: run the original harness as a reference, propose cross-domain and domain-specific edits, run each on the same task, and compose a new harness from the behavioral differences, with no ground-truth answer needed. Harnessing level: store adopted harnesses and evidence-backed lessons in a library for later tasks.

The paper reports that one step improves GLM-5 and Gemini 3.5 Flash over their base harnesses by up to 11.5 points on WorkBuddy-Bench, 4.5 on JobBench and 11.8 percentage points on LAB's partial pass rate, with more gains from more iterations. The code supports OpenCode and Codex executors and has a local demo that needs no API key. The repo has about 3,100 stars; X discussion picked up over the past two days.

【Verifiable】

GitHub: https://github.com/google-research/harness2

arXiv: https://arxiv.org/abs/2610.14224

【Our read】

1. A research paper with open code, not a Google product launch; online since October 9, so this is a catch-up entry.

2. Gains are self-reported and scored by an LLM judge (WorkBuddy setup requires a Gemini judge); no independent replication yet.

3. Like DeepSeek's open-source dsh and the Apple/EPFL single-session agent study, it shifts competition toward self-adjusting harnesses. The cost: each task runs several extra times, multiplying compute and API spend.