Created by a team of researchers from Chinese and US universities, MetaGPT is a new LLM-based meta programming framework aiming to enable collaboration in multi-agent systems by leveraging human procedural knowledge to enhance robustness, reduce errors, and engineer software solutions for complex tasks.
In this work, we present MetaGPT, a meta programming technology that utilizes SOPs to coordinate LLM-based multi-agent systems. SOPs serve as our strategy for organizing the collaboration of multi-agents which enhances the efficiency of their cooperative efforts.
MetaGPT takes a one-line requirement and builds user stories, competitive analysis, requirements, data structures, APIs, and other documents. It achieves that by replicating the structure of a software company. The researchers showed a demo where they ask MetaGPT to create a CLI blackjack game and have it generate all required artifacts including requirements, tests, and a working Python implementation of the game.
The central idea behind MetaGPT is encoding Standardized Operating Procedures (SOPs) into prompts to replicate efficient procedural knowledge required for collaborative tasks. The Agile Manifesto as well as other methods to distribute tasks and responsibilities across a team are examples of SOPs in the software field, say the researchers, including the definition of desired output, such as high-quality requirements documents, design artifacts, flowcharts, and interface specifications.
Similarly, SOPs use role-based action specifications and share an environment that enables them to actively observe one another and retrieve relevant information, which is a more efficient approach compared to passively receiving data through dialogue, the researchers say. For example, MetaGPT organizes its agents in product managers, architects, project managers, and engineers.
The image above shows the two main layers that define MetaGPT architecture: the Foundational Components Layer, and the Collaboration Layer. While the former allows agents to carry through their operations, the latter facilitates agent coordination through knowledge sharing and workflow encapsulation.
MetaGPT is not the only framework for meta programming of collaborative AI agents that aim to enable collaboration through some kind of task decomposition. Existing frameworks include AutoGPT, LangChain, and AgentVerse. According to MetaGPT team, their framework can handle higher levels of software complexity, with a 100% task completion rate.
MetaGPT is far from being a perfect system for AI collaboration and still requires work to deal with the hallucinatory tendency of LLM systems, which may lead, for example, to MetaGPT referencing non-existent resource files, or invoking undefined or not imported classes or variables.
If you are interested in the full details, do not miss the official paper, which includes a thorough description of the framework design, a detailed analysis of achieved results, and a comparison with alternative approaches.