No free wins

No Free Wins is an AI agent team that makes Python code faster and only counts a speedup it can prove. You give it a repo or GitHub URL and a goal such as “make top_words faster”. An analyst agent reads the code, picks the target function, writes a test-input generator and proposes a plan. In guided mode it asks questions or explains why a request can’t be optimized safely. In one-click autonomous mode it records its assumptions, repairs its own proposal if setup rejects it, and re-plans with lessons from a failed run, all within a single spend cap. A planner and two Nemotron workers with different strategies then write candidates, and an independent verifier reviews the selected one, bound to its exact source hash. Completion is decided by tests, not by the models. A deterministic harness runs every candidate in a locked-down Daytona sandbox, checks it against the original function on 2,000 hidden cases, and times it in 30 matched pairs. A run counts as done only when the median speedup and the 95% confidence interval’s lower bound both clear 3×. It then writes an audited receipt and a hash-chained ledger. Retries, model fallback, bounded rounds, budgets and deadlines keep every run autonomous and finite. Our official frozen session reached 3.54× (95% CI 3.49–3.62×), and analyst-driven runs on a user repo reached 8.03× and 8.10×.