Harness²
Open source framework (Apache 2.0) from Google Research that improves the agent without touching the model: it rewrites the harness around it. Eight components can be optimized: system prompts, guardrails, memory, skills, tool guidance, scripts, subagents and plugins. Each variant is compared with the others, the best is kept, the loop restarts. Code published on October 6, 2026, paper on arXiv on October 9. The repository states it is not an officially supported Google product.
Strengths
- Improves the agent without touching the model: same model, optimized surroundings
- Local demo with no model key, the improvement loop can be watched live
- Apache 2.0 licence on the core repository, code and paper published together
Limitations
- Gains reported by the authors only, not independently reproduced
- High inference cost and imperfect candidate selection, noted by press coverage
- An agent that rewrites its own guardrails: try it only in a sandbox
Best for
- Understanding automatic harness optimization before it reaches everyday tools