MetaAgent: Toward Self-Evolving Agent via Tool Meta-Learning

Source: arXiv Authors: Hongjin Qian, Zheng Liu Date: 2025-08-01 Primary category: cs.AI All categories: cs.AI, cs.CL, cs.IR

Abstract

MetaAgent begins with a minimal workflow and improves by organizing tool-use history, distilling lessons, and building in-house tools plus a persistent knowledge base. Architecturally it is interesting because the evolutionary surface is tool strategy and internal capability growth, not only workflow topology.