Automated Design of Agentic Systems

Source: arXiv Authors: Shengran Hu, Cong Lu, Jeff Clune Date: 2024-08-15 Primary category: cs.AI All categories: cs.AI

Abstract

This paper argues for a research area devoted to automatically designing agentic systems instead of hand-authoring them indefinitely. Its Meta Agent Search loop has agents program new agents in code, maintain an archive of prior discoveries, and iteratively search for better designs across domains. The paper is broader than workflow search alone, but it provides the cleanest statement of the meta-level thesis: prompts, tools, and workflows should all be learnable design surfaces.