ReASearch: Agent-Driven Optimization for Prompts, Programs, and ML Workflows

Traditional systems for optimizing prompts, programs, and machine learning workflows often rely on explicit outer-loop controllers like evolutionary search or textual gradients. This research explores a different paradigm, asking how much of this complex search policy can be internalized directly by a single, tool-using agent. The result is ReASearch, a unified framework for reasoning-driven optimization. In ReASearch, the agent itself autonomously decides what to evaluate, how to diagnose failures, which edits to make, and when to verify or restart. Instead of merely generating proposals based on heuristics, the agent actively analyzes outcomes, manages its budget, and refines its strategy over extended periods using persistent memory. By employing a shared agent loop with domain-specific tools, ReASearch can optimize prompts, programs, and ML workflows using the same underlying scaffold. Across 14 diverse tasks, ReASearch consistently outperforms or matches specialized optimization systems, achieving significant gains and even discovering solutions superior to previous human best-known results.