Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

EvoAgent: Towards Automatic Multi-Agent Generation via Evolutionary Algorithms

Siyu Yuan, Kaitao Song, Jiangjie Chen, Xu Tan, Dongsheng Li, Deqing Yang

TL;DR

EvoAgent reframes agent generation as an evolutionary process to automatically extend specialized LLM-based agents into multi-agent systems without human-designed scaffolds. It introduces a four-stage pipeline—initialization, crossover and mutation, selection, and results update—operating on agent settings and prompts to produce diverse, high-performing agents. Across NLP, multimodal benchmarks, interactive scientific solving, and real-world planning, EvoAgent demonstrates significant performance gains over strong baselines and existing auto-generation frameworks, while also offering insightful ablations on population size and iteration effects. The approach is framework-agnostic, scalable, and highlights practical considerations such as token costs and safety, suggesting EvoAgent as a versatile tool for constructing collaborative AI systems.

Abstract

The rise of powerful large language models (LLMs) has spurred a new trend in building LLM-based autonomous agents for solving complex tasks, especially multi-agent systems. Despite the remarkable progress, we notice that existing works are heavily dependent on human-designed frameworks, which greatly limits the functional scope and scalability of agent systems. How to automatically extend the specialized agent to multi-agent systems to improve task-solving capability still remains a significant challenge. In this paper, we introduce EvoAgent, a generic method to automatically extend specialized agents to multi-agent systems via the evolutionary algorithm, thereby improving the effectiveness of LLM-based agents in solving tasks. Specifically, we consider the existing agent frameworks as the initial individual and then apply a series of evolutionary operators (e.g., mutation, crossover, selection, etc.) to generate multiple agents with diverse settings. Experimental results across various tasks show that EvoAgent can significantly enhance the task-solving capability of LLM-based agents, and can be generalized to any LLM-based agent framework to extend them into multi-agent systems. Resources are available at https://evo-agent.github.io/.

EvoAgent: Towards Automatic Multi-Agent Generation via Evolutionary Algorithms

TL;DR

EvoAgent reframes agent generation as an evolutionary process to automatically extend specialized LLM-based agents into multi-agent systems without human-designed scaffolds. It introduces a four-stage pipeline—initialization, crossover and mutation, selection, and results update—operating on agent settings and prompts to produce diverse, high-performing agents. Across NLP, multimodal benchmarks, interactive scientific solving, and real-world planning, EvoAgent demonstrates significant performance gains over strong baselines and existing auto-generation frameworks, while also offering insightful ablations on population size and iteration effects. The approach is framework-agnostic, scalable, and highlights practical considerations such as token costs and safety, suggesting EvoAgent as a versatile tool for constructing collaborative AI systems.

Abstract

The rise of powerful large language models (LLMs) has spurred a new trend in building LLM-based autonomous agents for solving complex tasks, especially multi-agent systems. Despite the remarkable progress, we notice that existing works are heavily dependent on human-designed frameworks, which greatly limits the functional scope and scalability of agent systems. How to automatically extend the specialized agent to multi-agent systems to improve task-solving capability still remains a significant challenge. In this paper, we introduce EvoAgent, a generic method to automatically extend specialized agents to multi-agent systems via the evolutionary algorithm, thereby improving the effectiveness of LLM-based agents in solving tasks. Specifically, we consider the existing agent frameworks as the initial individual and then apply a series of evolutionary operators (e.g., mutation, crossover, selection, etc.) to generate multiple agents with diverse settings. Experimental results across various tasks show that EvoAgent can significantly enhance the task-solving capability of LLM-based agents, and can be generalized to any LLM-based agent framework to extend them into multi-agent systems. Resources are available at https://evo-agent.github.io/.
Paper Structure (48 sections, 4 figures, 7 tables)

This paper contains 48 sections, 4 figures, 7 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. LLM-based Autonomous Agents
  4. Agent Generation
  5. Method
  6. Preliminary
  7. EvoAgent
  8. STEP 1: Initialization
  9. STEP 2: Crossover & Mutation
  10. STEP 3: Selection
  11. STEP 4: Results Update
  12. Experiment
  13. NLP and Multi-Modal Tasks
  14. Benchmarks
  15. Baselines
  16. ...and 33 more sections

Figures (4)

  • Figure 1: The illustration of EvoAgent. With the generated multiple specialized agents, EvoAgent can generate a better travel plan to meet user preferences. For EA operators, Crossover can improve the results of parent agents by adjusting existing details (e.g., the information marked as blue). Mutation can introduce new variations to refine the results of parent agents by generating child agents with new characteristics (e.g., the information marked as red).
  • Figure 2: Overall results of GPT-4V and Gemini-Pro with different methods on the MMMU validation set. We also compare the performance of GPT-4V and Gemini-Pro across three difficulty levels.
  • Figure 3: The adaption of EvoAgent on MetaGPT framework. With the EA, we can extend the original role in the debate scenario to different specialized agents to enrich the opinions.
  • Figure 4: The adaption of EvoAgent on Camel and AutoGen frameworks.