ReASearch: Agents Autonomously Optimize Prompts and ML Workflows via Reasoning
Junbo Li · hf · 2026-08-10
Current systems for optimizing prompts, programs, and ML workflows typically rely on explicit outer-loop controllers like evolutionary search. Researchers proposed ReASearch, a framework exploring the internalization of search policies within a single tool-using agent.
- Core Mechanism: The agent autonomously decides what to evaluate, diagnoses failures, makes edits, and verifies or restarts. It actively analyzes outcomes, allocates budget, and refines strategies over long horizons via persistent memory.
- Unified Scaffold: Using a shared agent loop and domain-specific tools, ReASearch instantiates the exact same scaffold to optimize prompts, programs, and ML workflows.
- Significant Results: Across 14 diverse tasks, it performs better than specialized optimization systems, achieving gains of 2% to 40% over strong baselines, occasionally discovering solutions that improve upon prior human best-known results. Complex search behaviors emerge naturally from the agent's reasoning process.
More from coding & agent
- Radar: Open-Source Kubernetes UI with Built-in MCP Server for AI Agents — tom_doerr · 2026-08-10
- When AI Becomes a Co-Contributor: Devs Share Relatable Meme — cneuralnetwork · 2026-08-10
- Open-Sourcing a 1MB 3D Web Experience Built with Claude Opus — moeinteractive · 2026-08-10
- 10 Ways to Cut AI Agent Token Usage by 72% Without Losing Accuracy — Emotional_Collar_449 · 2026-08-10
- pbx-mcp: An Open-Source MCP Server for Secure AI Control of Asterisk — ictinnovations · 2026-08-10
- Vjeux Fixes Agent Scaling Bottleneck: Replacing Frequent Worktree Creation with Folder Pooling — Vjeux · 2026-08-10