Google opens Harness²: AI agent improves its own working environment

Google Research has released Harness², an Apache 2.0-licensed framework that recursively improves an AI agent's environment—prompt, guardrails, memory, sub-agents, skills, scripts, and plugins—using a Propose–Probe–Compose scheme, achieving up to an 11.5-point gain on WorkBuddy-Bench in a single step with Gemini 3.5 Flash and GLM-5.

Compiled by Sergey KostenchukPublished 2026-10-18Updated 2026-10-19

2026-10-18 Coding Google

🛠 Google opens Harness²: AI agent improves its own working environment

Google Research has released Harness², an open-source framework (Apache 2.0): the agent recursively improves its own harness—the environment surrounding the model: system prompt, guardrails, memory, sub-agents, skills, scripts, and plugins. The Propose–Probe–Compose mechanism tests edit variants on a single task and assembles a configuration from behavioral contrasts, without a ready-made answer. On Gemini 3.5 Flash and GLM-5, a single step yielded up to an 11.5-point gain on WorkBuddy-Bench.

🌍 Agent productivity grows without model fine-tuning—adaptation to a new task is cheaper. Contrastive improvement does not guarantee correctness; independent verification is required for legal and safety-critical tasks.

👤 The framework runs locally (Linux, Python 3.12, uv), the demo works without API keys, and the method is described in arXiv:2610.14224.

Source 1: https://github.com/google-research/harness2 Source 2: https://arxiv.org/abs/2610.14224

Sources