[Paper Review] Multi-Agent Collaboration: Harnessing the Power of Intelligent LLM Agents

Yashar Talebirad, Amirhossein Nadiri|arXiv (Cornell University)|Jun 5, 2023

Artificial Intelligence in LawSocial Sciences50 citations

TL;DR

The paper proposes a general multi-agent framework for enhancing LLM capabilities by coordinating multiple intelligent generative agents (IGAs) and plugins, and demonstrates its applicability to Auto-GPT, BabyAGI, and Gorilla-like systems.

ABSTRACT

In this paper, we present a novel framework for enhancing the capabilities of large language models (LLMs) by leveraging the power of multi-agent systems. Our framework introduces a collaborative environment where multiple intelligent agent components, each with distinctive attributes and roles, work together to handle complex tasks more efficiently and effectively. We demonstrate the practicality and versatility of our framework through case studies in artificial general intelligence (AGI), specifically focusing on the Auto-GPT and BabyAGI models. We also examine the "Gorilla" model, which integrates external APIs into the LLM. Our framework addresses limitations and challenges such as looping issues, security risks, scalability, system evaluation, and ethical considerations. By modeling various domains such as courtroom simulations and software development scenarios, we showcase the potential applications and benefits of our proposed multi-agent system. Our framework provides an avenue for advancing the capabilities and performance of LLMs through collaboration and knowledge exchange among intelligent agents.

Motivation & Objective

• Introduce a general framework for multi-agent collaboration among IGAs to push toward more capable AI systems and potential AGI.
• Enable adaptive, dynamic system configuration with agent addition/removal to handle diverse tasks.
• Facilitate knowledge exchange and error mitigation through inter-agent and self-feedback mechanisms.
• Provide a structured abstraction that can model existing AGI frameworks and their components (agents, plugins, memories, APIs).
• Highlight practical use cases (courtroom simulations, software development) and discuss challenges and opportunities.

Proposed method

• Model the multi-agent environment as a black-box graph G(V,E) of IGAs and plugins.
• Represent each agent as a tuple A_i=(L_i,R_i,S_i,C_i,H_i) detailing model, role, state, ability to create agents, and halt permissions.
• Represent each plugin as a tuple P_j=(F_j,C_j,U_j) with functionalities, configurations, and usage constraints.
• Specify message exchanges as m=(S_m,A_m,D_m) between agents via edges in E, enabling tasks, reports, and requests.
• Describe system design steps: assign agent roles, connect agents to plugins, enable agent–agent communication, and manage creation/halt permissions.
• Introduce dynamic agent addition, where capable agents can instantiate new agents with inherited properties and connections.
• Incorporate feedback mechanisms: inter-agent feedback and self-feedback to improve strategies and performance.
• Define an Oracle agent that is stateless/memory-less for isolated inputs and utility to other agents (e.g., feedback).
• Incorporate a halting/supervision mechanism, possibly via a Supervisor Agent, to manage and stop undesired agent behaviors.
• Propose autonomous system design where an IGA can design or refine the entire system’s structure and roles.

Experimental results

Research questions

• RQ1 How can a framework of multiple IGAs improve task performance compared to isolated LLMs?
• RQ2 How should agents, plugins, and communication be modeled to enable efficient collaboration and scalability?
• RQ3 What mechanisms (feedback, supervision, halting) are effective for controlling multi-agent dynamics?
• RQ4 How can existing AGI frameworks (Auto-GPT, BabyAGI, Gorilla) be mapped into the proposed framework to improve design and operation?
• RQ5 What are the practical applications and challenges of deploying such a framework in domains like law and software development?

Key findings

• The framework is demonstrated as applicable to Auto-GPT, BabyAGI, and Gorilla-style models by mapping their components to agents and plugins.
• Dynamic agent creation and supervisory mechanisms can help manage workload, mitigate loops, and improve collaboration.
• Feedback mechanisms (inter-agent and self-feedback) are proposed to enable learning and adaptation without constant human input.
• Case studies illustrate modeling complex domains (court simulations and software development) within the multi-agent abstraction.
• The framework identifies limitations (looping, security, scalability, evaluation, ethics) and discusses potential improvements and safeguards.
• A specialized Supervisor Agent could monitor progress and trigger halts to prevent stagnation or deviation.