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

MegaAgent is a framework designed to manage autonomous parallel $O(\log n)$ cooperation in large-scale LLM-powered multi-agent systems, enabling dynamic agent generation, task auto-splitting, and enhanced communication without relying on predefined operating procedures (like human-written prompts, which is impossible for large number of agents).

Framework

Code Overview

This repository contains two parts: the latest version of MegaAgent at ., and examples at examples/.

To run the latest version, you can add your key and change the prompt in config.py, and run main.py. The output will be located at files/. The default prompt is used for automatically generating a Gobang game.

Some experiments are shown in examples/ using an older version of MegaAgent. You can use the same prompt while substituting other files with the latest version.

Backbone / API interface

The main MegaAgent code (.) and the TravelPlanner example talk to the model through the modern OpenAI Python SDK tool-use interface (tools / tool_calls / role:"tool"). The shared transport lives in a single file, llm_core.py, which their llm.py delegates to; all framework logic, prompts, and tool schemas are unchanged from the original design. Configure the backbone in config.py:

api_key = 'YOUR_KEY'
model = "gpt-5.6-sol"
base_url = 'https://your-endpoint/v1'
reasoning_effort = 'xhigh'   # optional; sent only when set

llm_core.py streams every request internally (reassembling one complete response), sends no temperature and never caps max_tokens (so long reasoning is never truncated), and sanitizes histories to the strict tool-use protocol. Install dependencies with pip install -r requirements.txt.

The other examples under examples/ still use the legacy functions/function_call interface with url in their config.py.

Experimental Results

RQ1: Quantitative experiments using gpt-4o as backbone

Model	MBPP	HumanEval	MATH	GSM-8k
MetaGPT	81.7%	82.3%	N/A	N/A
Camel	78.1%	57.9%	22.3%	45.6%
AgentVerse	82.4%	89.0%	54.5%	81.2%
AutoGen	85.3%	85.9%	69.5%	87.8%
MegaAgent	92.2%	93.3%	69.0%	93.0%

RQ2: Gobang Game Codebase with multiple code files

RQ3: Large-Scale National Policy Simulation involving 590+ dynamically generated parallel agents

We also evaluated MegaAgent on TravelPlanner (validation set, sole-planning mode). The submission file (merged_plans.jsonl) is included in examples/travel planner.

Metric	GPT-4o	GPT-5.6
Delivery Rate	100.0%	100.0%
Commonsense Constraint Micro Pass Rate	81.88%	97.64%
Commonsense Constraint Macro Pass Rate	27.22%	84.44%
Hard Constraint Micro Pass Rate	40.48%	87.14%
Hard Constraint Macro Pass Rate	23.89%	83.33%
Final Pass Rate	10.0%	76.67%

The GPT-5.6 column uses gpt-5.6-sol with reasoning_effort=xhigh through the tool-use interface described above.

Licenses

This repository is under license CC BY 4.0.

Acknowledgement

We would like to thank Xinyi Zhang for editing the poster.

Citation

If you find this repository useful, please cite our paper:

@inproceedings{wang-etal-2025-megaagent,
    title = "{M}ega{A}gent: A Large-Scale Autonomous {LLM}-based Multi-Agent System Without Predefined {SOP}s",
    author = "Wang, Qian  and
      Wang, Tianyu  and
      Tang, Zhenheng  and
      Li, Qinbin  and
      Chen, Nuo  and
      Liang, Jingsheng  and
      He, Bingsheng",
    editor = "Che, Wanxiang  and
      Nabende, Joyce  and
      Shutova, Ekaterina  and
      Pilehvar, Mohammad Taher",
    booktitle = "Findings of the Association for Computational Linguistics: ACL 2025",
    month = jul,
    year = "2025",
    address = "Vienna, Austria",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2025.findings-acl.259/",
    doi = "10.18653/v1/2025.findings-acl.259",
    pages = "4998--5036",
    ISBN = "979-8-89176-256-5",
    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 \textit{MegaAgent}, a large-scale autonomous LLM-based multi-agent system. \textit{MegaAgent} generates agents based on task complexity and enables dynamic task decomposition, parallel execution, efficient communication, and comprehensive system monitoring of agents. In evaluations, \textit{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, \textit{MegaAgent} demonstrates exceptional scalability and autonomy, setting a foundation for advancing true autonomy in MAS."
}