Multi-agent frameworks

LLM-based multi-agent frameworks are software platforms that enable multiple autonomous AI agents, built on large language models (LLMs), to interact with each other to collaboratively solve complex problems. Such systems emphasize a diversity of agent profiles, their communication, and collective decision-making. This approach leverages the "collective intelligence" of a group, where each agent performs a specialized role, and the exchange of messages between them simulates human collaboration.

This allows for the modeling of complex real-world scenarios and the solving of problems that exceed the capabilities of a single intelligent agent. Multi-agent LLM systems have already demonstrated successful results in areas such as software development, social simulations, economic games, and modeling political discussions.

Key Frameworks and Approaches

The development of multi-agent systems has led to the emergence of several open-source frameworks that facilitate their creation and research.

MetaGPT (2023)

One of the first open-source frameworks, focused on collaborative work following an assembly line principle. MetaGPT incorporates standard operating procedures (SOPs) into the system and assigns each agent a specific role (e.g., product manager, engineer, QA tester). This approach allows for the decomposition of a complex task into sub-tasks, distributing them among specialized agents, which reduces chaos and the risk of hallucinations.

CAMEL (2023)

The CAMEL (Communicative Agents for "Mind" Exploration) framework focuses on the autonomous interaction of agents through dialogue. It proposes an inception prompting technique to coordinate conversations between LLM agents, guiding them toward a common goal. Each agent is assigned a role and a context, after which the agents communicate in natural language, gradually developing a joint solution. CAMEL has proven effective in scenarios requiring cooperation without direct human intervention.

AutoGen (2023)

A versatile and customizable framework from Microsoft researchers, designed for creating complex applications based on multi-LLM communication. AutoGen allows the logic of agent interaction to be programmed using both code and natural language. It supports integration with external tools and APIs, making it suitable for a wide range of tasks—from software development to creating conversational systems.

AgentVerse (2023)

An open platform developed by the OpenBMB community for studying dynamic collaboration and emergent agent behaviors. AgentVerse provides two modes of operation:

1. Task-solving: Multiple LLM agents form a team to complete a complex task (e.g., collaborative software development).
2. Simulation: Allows the user to define a virtual environment and observe the interactions of agents (e.g., simulating a classroom or the prisoner's dilemma).

The platform emphasizes the importance of a standardized environment and communication protocols for managed communication.

CrewAI (2024)

A framework focused on integrating LLM agents into business processes and data analysis. CrewAI implements the AI-Based Agents Workflow (AgWf) concept, where agents execute steps described as text instructions and can use external tools (Python classes/functions). This enables the automation of complex analytical scenarios by combining the flexibility of LLMs with deterministic code.

LangGraph (2024)

An experimental framework that uses graph structures to represent state and context in dialogues with LLMs. The key feature of LangGraph is its support for cyclical workflows. This allows agents to exchange data through a shared knowledge graph, iteratively search for information, assess its reliability, and correct their responses, which is particularly useful in question-answering (QA) tasks with an augmented knowledge base.

Other Projects

The experimental projects AutoGPT and AgentGPT also gained widespread attention, demonstrating the potential of fully autonomous AI agents capable of independently setting goals, performing web searches, executing code, and managing files. Although these projects were not peer-reviewed, they highlighted the importance of planning, memory, and tool-use components for building truly autonomous agents.

Applications of Multi-Agent Systems

• Software Development Automation: Groups of LLM agents act as managers, programmers, and testers, collaboratively planning and implementing software projects. The ChatDev study showed that a team of four agents could create a simple application in minutes, engaging in dialogue at all stages from task definition to testing.
• Intelligent Assistants: Enterprise products like Microsoft 365 Copilot and IBM Watsonx Orchestrate use multiple agents to perform complex tasks, where one agent processes a request, another retrieves facts from a database, and a third compiles a report.
• Scientific Research: Agents are used to generate and critique hypotheses. In approaches like Guided Debate or Self-Refine, one agent proposes a solution, while another evaluates and corrects it, which helps reduce errors.
• Social Modeling and Virtual Worlds: In the landmark project Generative Agents, dozens of LLM agents endowed with personalities and memory simulated life in a small virtual town, demonstrating believable social interaction. Such simulations could find applications in gaming (for creating realistic NPCs), education, and the social sciences.

Challenges and Future Outlook

Despite their successes, multi-agent systems face several significant challenges:

• Hallucinations and Cascading Errors: An error made by one agent can be passed down the chain to others, who then use it as the basis for their reasoning, leading to skewed results for the entire group.
• Scalability and Resource Intensity: Each LLM-based agent requires significant computational resources. Ensuring the simultaneous operation of dozens of agents is a complex technical challenge.
• Coordination and Management: As the number of agents grows, the risk of chaos increases. Sophisticated "orchestration" mechanisms are needed to manage their interactions.
• Evaluation and Testing: There is a lack of widely accepted benchmarks for objectively comparing different multi-agent frameworks.

In the future, the emergence of more efficient and secure multimodal systems is expected, where agents can exchange not only text but also images and other data. The implementation of reinforcement learning could teach groups of agents to coordinate better over time, achieving a "collective intelligence" effect. Ultimately, multi-agent frameworks represent a step toward creating more flexible and powerful AI systems, where multiple specialized "minds" work together.

Links

• Survey Paper: Large Language Model based Multi-Agents (2024)
• Survey Paper: LLMs Working in Harmony (2025)

Further Reading

• Yao, S. et al. (2022). ReAct: Synergizing Reasoning and Acting in Language Models. arXiv:2210.03629.
• Hong, S. et al. (2023). MetaGPT: Meta Programming for a Multi-Agent Collaborative Framework. arXiv:2308.00352.
• Li, G. et al. (2023). CAMEL: Communicative Agents for "Mind" Exploration of Large Language Model Society. arXiv:2303.17760.
• Wu, Q. et al. (2023). AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation. arXiv:2308.08155.
• Liu, X. et al. (2023). AgentBench: Evaluating LLMs as Agents. arXiv:2308.03688.
• Yao, S. et al. (2023). Tree of Thoughts: Deliberate Problem Solving with Large Language Models. arXiv:2305.10601.
• Shinn, N. et al. (2023). Reflexion: Language Agents with Verbal Reinforcement Learning. arXiv:2303.11366.
• Chen, W. et al. (2023). AgentVerse: Facilitating Multi-Agent Collaboration and Exploring Emergent Behaviors. arXiv:2308.10848.
• Park, J. S. et al. (2023). Generative Agents: Interactive Simulacra of Human Behavior. arXiv:2304.03442.
• Guo, T. et al. (2024). Large Language Model Based Multi-Agents: A Survey of Progress and Challenges. arXiv:2402.01680.
• Chen, Q. et al. (2024). ChatDev: Communicative Agents for Software Development. arXiv:2307.07924.
• Duan, Z.; Wang, J. (2024). Exploration of LLM Multi-Agent Application Implementation Based on LangGraph + CrewAI. arXiv:2411.18241.
• Aratchige, R. M.; Ilmini, W. M. K. S. (2025). LLMs Working in Harmony: A Survey on the Technological Aspects of Building Effective LLM-Based Multi-Agent Systems. arXiv:2504.01963.