Elevating long-horizon agentic tasks with orchestrated Test-Time Compute

AI21 Maestro is a general-purpose agentic framework that improves model performance on complex tasks by optimizing test-time compute (TTC) allocation. It separates decision-making from the LLM itself, allowing for adaptive resource management and better control over long-horizon tasks. When applied to the SWE-bench benchmark, Maestro significantly boosts the performance of models like GPT-5 and GPT-5-mini. Key components contributing to this improvement include horizontal scaling, which runs multiple cheaper models in parallel, and the use of structured plans to guide execution.