The recurring challenges in AI agent development, how to find better prompts, evaluate improvements reliably, and maintain diversity in search, are the same fundamental problems that reinforcement learning and evolutionary search have grappled with for decades. Gear, an open-source framework from a research team, directly addresses these issues by separating search decisions from execution and evaluation, treating them as distinct, optimizable components. This approach is a practical step toward making recursive self-improvement (RSI) more systematic, and it mirrors the logic we explored in How Your Agent's Tools Can Make Smarter Decisions, where decision models themselves become tools for better reasoning. The team's goal is to create a "scikit-learn for RSI," enabling easier comparison of methods like GEPA, RRSI, and Harness².
The results from their initial tests with GPT 5.6 Luna are telling. On a standard benchmark, Terminal-Bench 2.1, the base task pass rate of 52.8% jumped to 62.9% after Gear optimized the agent's prompts and workflow. The improvement was even more striking on AutomationBench Marketing, where the pass rate rose from 27.0% to 40.0%. At maximum reasoning effort, Luna with Gear's optimized harness hit 88.88% objective completion, compared to 84.08% for a competing system using Astra + Codex. This advantage came from partial completion rather than strict pass rates, which were 53% versus 57%. That distinction matters: it suggests the framework excels at helping agents get further along in complex tasks, even if they don't always finish perfectly. Boost Your AI Agent's Performance Without Changing the Model made a similar case, that system design often matters more than raw model capability, and Gear's results reinforce that point with hard numbers.
Perhaps the most striking finding is the cost difference. Using the same optimized configuration at maximum effort, Luna's estimated API cost averaged $0.0568 per task, while Astra's came in at $1.1858, a 95.21% reduction. That's not a marginal efficiency gain; it's a fundamental shift in what becomes economically feasible. If you can run dozens of task attempts for the price of one, you can afford to iterate more aggressively, test more hypotheses, and fail more often without breaking your budget. This aligns with the broader trend Anthropic revealed in their new tracking tools, where transparency around spending is becoming a priority for serious AI deployment.
The open question that remains is diversity preservation under limited budgets. Gear's current GEPA variant uses shared failure diagnosis and staged evaluation, but the team is asking for feedback on how to prevent premature convergence, the trap where the optimizer finds a local optimum and stops exploring. If they can solve that, they'll have built something genuinely useful: a framework that makes RSI not just possible, but practical and cost-effective. The specific detail to watch is how their candidate evaluation handles trade-offs between exploration and exploitation when each API call costs real money.