Budget-Aware Best-of-N for SWE Agents

AI21 research on making best-of-N sampling affordable for software agents: predictive "Resolve-Now / Resolve-Later" classifiers drive cascading execution (up to 44% compute reduction at matched resolve rate) and parallel execution with mid-air termination (up to 25% speedup), evaluated on a 123-issue SWE-rebench slice. Blog-only — no paper or code released.