Gear: Scikit-learn for RSI

Gear is an open-source framework for recursive self-improvement of AI agent harnesses (prompts, tools, workflows, and context management). Built toward a scikit-learn for RSI, its modular architecture lets developers implement, compare, and extend optimization algorithms. With Gear, GPT 5.6 Luna reaches GPT 6 Astra-level performance: 88.9% vs. Astra + Codex’s 84.1% on AutomationBench’s public Marketing tasks.