EVOMIND: A Framework for Self-Evolving Intelligent Agents

This paper proposes a self-evolving multi-agent artificial intelligence framework that aims to tackle difficult open-ended problems through autonomous collaboration with little human interaction. Unlike conventional static systems, this modular framework allows Large Language Model (LLM)-based agents to adapt their operational methods through continuous feedback adaptation. The proposed framework is organized around a customized hierarchy of an agent manager, executor, evaluator, and refiner, along with a shared memory module and an iterative feedback process. In essence, the framework breaks down user-specified tasks into individual subtasks, assigns them to dedicated agents, and combines their results. Most importantly, the framework examines performance feedback to automatically optimize prompts, workflows, and role assignments for the next iteration. The experimental outcome clearly shows that this self-evolving framework performs much better than static single-agent baselines in terms of task quality and temporal consistency. Moreover, the modular framework is amenable to the seamless incorporation of external tools and new roles of agents across various application domains. This work implies that the integration of multi-agent orchestration and automated evolution is a critical milestone towards the development of robust, lifelong learning artificial intelligence systems that can efficiently tackle real-world problems.