After reading a bunch of RSI papers, I keep coming back to familiar problems from RL and evolutionary search:

1. Sampling: How do we find promising changes efficiently?
2. Evaluation: How do we know a new variant actually improves the agent?
3. Diversity: How do we avoid converging too early?

That motivated us to build Gear, an open-source, meta-learning-inspired framework. A task agent produces execution traces; a meta-agent uses those traces and evaluation feedback to improve prompts, tools, workflows, and context management.

Gear separates search decisions from execution and evaluation. Our goal is a scikit-learn for RSI, making approaches such as GEPA, RRSI, and Harness² easier to implement and compare. We currently have a working GEPA variant with shared failure diagnosis and staged evaluation.

With GPT 5.6 Luna held at medium reasoning effort, we observed these task pass rates:

Benchmark	Base DSH	Gear-optimized DSH
AutomationBench Marketing	27.0%	40.0%
AutomationBench Support	17.0%	25.0%
Terminal-Bench 2.1	52.8%	62.9%

At max effort on Marketing, Luna with the optimized harness reached 88.88% objective completion, versus 84.08% for Astra + Codex. Strict task pass rates were 53% versus 57%, so the advantage was in partial completion.

Separately, with both models using the same optimized DSH at max effort, Luna’s estimated API cost averaged $0.0568 per task versus $1.1858 for Astra—95.21% lower, excluding optimization and infrastructure costs.

We’d welcome feedback on candidate evaluation and diversity preservation under limited budgets.

Code · Paper