Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

MegaAgent: A Large-Scale Autonomous LLM-based Multi-Agent System Without Predefined SOPs

Qian Wang, Tianyu Wang, Zhenheng Tang, Qinbin Li, Nuo Chen, Jingsheng Liang, Bingsheng He

TL;DR

MegaAgent addresses the scalability and autonomy gap in LLM-based multi-agent systems by introducing a hierarchical, OS-inspired framework that autonomously decomposes tasks, forms dynamic agent groups, executes tasks in parallel, and monitors results without predefined SOPs. The system combines a multi-level task management stack (Boss-Agent decomposition, dynamic group formation) with a robust hierarchical monitoring and a producer–consumer message-queue, plus Git-based file management and vector-memory for persistent context. Empirical results show MegaAgent outperforms baselines on standard benchmarks, and achieves full Gobang game development within 800 seconds and national policy generation with 590 agents in 2,991 seconds, supported by ablations and cost analyses that underscore the necessity of hierarchy, parallelism, and monitoring. The work demonstrates strong potential for real-world large-scale MAS applications, while acknowledging limitations in planning overhead, hallucinations, and API costs, and pointing to future directions in memory, token efficiency, and cheaper LLM integration.

Abstract

LLM-based multi-agent systems (MAS) have shown promise in tackling complex tasks. However, existing solutions often suffer from limited agent coordination and heavy reliance on predefined Standard Operating Procedures (SOPs), which demand extensive human input. To address these limitations, we propose MegaAgent, a large-scale autonomous LLM-based multi-agent system. MegaAgent generates agents based on task complexity and enables dynamic task decomposition, parallel execution, efficient communication, and comprehensive system monitoring of agents. In evaluations, MegaAgent demonstrates exceptional performance, successfully developing a Gobang game within 800 seconds and scaling up to 590 agents in a national policy simulation to generate multi-domain policies. It significantly outperforms existing systems, such as MetaGPT, in both task completion efficiency and scalability. By eliminating the need for predefined SOPs, MegaAgent demonstrates exceptional scalability and autonomy, setting a foundation for advancing true autonomy in MAS. Our code is available at https://github.com/Xtra-Computing/MegaAgent .

MegaAgent: A Large-Scale Autonomous LLM-based Multi-Agent System Without Predefined SOPs

TL;DR

MegaAgent addresses the scalability and autonomy gap in LLM-based multi-agent systems by introducing a hierarchical, OS-inspired framework that autonomously decomposes tasks, forms dynamic agent groups, executes tasks in parallel, and monitors results without predefined SOPs. The system combines a multi-level task management stack (Boss-Agent decomposition, dynamic group formation) with a robust hierarchical monitoring and a producer–consumer message-queue, plus Git-based file management and vector-memory for persistent context. Empirical results show MegaAgent outperforms baselines on standard benchmarks, and achieves full Gobang game development within 800 seconds and national policy generation with 590 agents in 2,991 seconds, supported by ablations and cost analyses that underscore the necessity of hierarchy, parallelism, and monitoring. The work demonstrates strong potential for real-world large-scale MAS applications, while acknowledging limitations in planning overhead, hallucinations, and API costs, and pointing to future directions in memory, token efficiency, and cheaper LLM integration.

Abstract

LLM-based multi-agent systems (MAS) have shown promise in tackling complex tasks. However, existing solutions often suffer from limited agent coordination and heavy reliance on predefined Standard Operating Procedures (SOPs), which demand extensive human input. To address these limitations, we propose MegaAgent, a large-scale autonomous LLM-based multi-agent system. MegaAgent generates agents based on task complexity and enables dynamic task decomposition, parallel execution, efficient communication, and comprehensive system monitoring of agents. In evaluations, MegaAgent demonstrates exceptional performance, successfully developing a Gobang game within 800 seconds and scaling up to 590 agents in a national policy simulation to generate multi-domain policies. It significantly outperforms existing systems, such as MetaGPT, in both task completion efficiency and scalability. By eliminating the need for predefined SOPs, MegaAgent demonstrates exceptional scalability and autonomy, setting a foundation for advancing true autonomy in MAS. Our code is available at https://github.com/Xtra-Computing/MegaAgent .
Paper Structure (77 sections, 40 figures, 17 tables, 2 algorithms)

This paper contains 77 sections, 40 figures, 17 tables, 2 algorithms.

Table of Contents

  1. Introduction
  2. MegaAgent Framework
  3. Overview
  4. Hierachical Task Management
  5. Multi-level Task Splitting
  6. Hierarchical Coordination Mechanism
  7. Message Queue Mechanism
  8. File Management
  9. Hierarchical Monitoring
  10. Multi-level Monitoring
  11. Failure Scenarios and Solutions
  12. Experiments
  13. RQ1: Standard Benchmarks
  14. Experiment Setup
  15. Experiment Results and Analysis
  16. ...and 62 more sections

Figures (40)

  • Figure 1: MegaAgent processes a user-provided meta-prompt by dividing it into distinct tasks, assigning each to a corresponding admin agent. Admin agents oversee their tasks, autonomously recruiting additional agents as needed to form task-specific groups that operate in parallel for efficient execution. These groups can further expand through sub-agent recruitment, creating a multi-level hierarchy. Admin agents supervise their groups to ensure task completion and output quality. Agents are classified into admin and ordinary agents: admin agents can communicate with one another, while ordinary agents interact only within their groups to optimize communication efficiency. Each agent is treated as a process whose state is either Idle, Processing, or Response. Agents access and manage external files in storage module using function calls, supporting seamless data retrieval and task execution.
  • Figure 2: Agent Communication Example: National Leader to Minister of Health
  • Figure 3: Monitoring Log Example: National Leader monitors the financial policy.
  • Figure 4: Gobang Game Meta Prompt
  • Figure 5: National Policy Generation Meta Prompt
  • ...and 35 more figures