MetaGPT: Is This the Best Multi-Agent Framework Yet?

The AI revolution is something we can’t underestimate. There are mixed reactions towards AI, with opposers saying that AI might become too powerful and humans will be unable to control it. There is also a group that believes AI will take their jobs. 

On the other hand, we have a faction who are already utilizing AI to become more productive and efficient. E-commerce platforms have been using artificial intelligence to study consumer behaviors.

Recently, we have seen AI products that can help users generate code with a few prompts, generate marketing copy, explain code, and generate images, videos, and thumbnails with high accuracy. 

New products are being created daily, and MetaGPT is one of the new entrants already making headlines. 

What is MetaGPT?

MetaGPT is a multi-agent framework that takes one-line inputs to produce APIs, user stories, data structures, competitive analysis, and more. GPT is the short form for Generative Pretrained Transformers. 

MetaGPT framework can behave as a product manager, software engineer, and architect. This framework can act as an entire software company with a single line of code with orchestrated SOPs. 

MetaGPT integrates with the human SOP process design. As a result, the LLM-based agent generates high-quality, diverse, structured documentation and designs. The design of MetaGPT makes it easy to engineer solutions for complex tasks and offers problem-solving capabilities that almost compare to human intelligence. 

This framework has two distinct layers; foundational components and collaboration layers. 

The foundational components layer has all the building blocks needed by individual agent operations. These building blocks are the Environment, Memory, Roles, Actions, and Tools. 

• The environment allows shared communications and workspaces 
• Roles define domain-specific workflows and skills
• Tools offer common utilities and services
• Actions take care of sub-tasks

The collaboration layer is built on the foundational components layer. This layer orchestrates different agents to resolve complex problems collaboratively. In this layer, SOPs break down tasks into small manageable components, assign them to suitable agents and supervise their execution to ensure all guidelines are followed. This layer also ensures that different agents share and exchange data, creating a shared knowledge base. 

Advantages of using MetaGPT

We already have hundreds of AI tools. What makes MetaGPT so special? These are the advantages of using MetaGPT;

• Automates software development process: MetaGPT automates most of the code generation process, giving developers more time to focus on strategic and creative tasks. 
• Pre-trained language understanding: MetaGPT is based on multiple GPTs already trained to perform different tasks. 
• Can generate creative programs: MetaGPT framework helps users generate fully-functional applications with a few commands. 
• Can enhance existing programs: This framework can study an existing program, add new features, or even do away with non-essential code. 
• Can facilitate communication: MetaGPT allows different team members to collaborate and communicate better as they work towards a common goal. 

How does MetaGPT work?

MetaGPT uses different agents (GPTs) to handle requests. Let us say if you want to create an app that transcribes music. MetaGPT will look at the prompt, identify the best-suited GPT, and assign them different tasks. This framework will then split the work into small chunks and assign it to different agents that are in communication during the entire process.

According to the data on its GitHub page, generating one example with design and analysis will cost you about $0.2 (in GPT-4 API fees). If you want a complete project, then you need about $2. This is how MetaGPT works under the hood;

• The user defines software requirements: The user writes the instructions based on the type of application they want to build. For instance, you can instruct this framework to build a hotel booking app with an admin dashboard and users. 
• MetaGPT identifies the right GPT: MetaGPT works with different GPTs. For instance, you can have different GPTs producing images while others generate code. 
• GPT works on the input: After identifying the right GPT, the framework will generate the output based on the pre-trained data. 
• User/s reviews the output: MetaGPT will process outputs based on the inputs. As a user, review them to determine if they suit your needs. 
• MetaGPT refines the output: If the outputs still lack some features, you can instruct the framework to refine them until satisfied. 

Getting Started with MetaGPT

Since you now understand how MetaGPT works, it is time to know how to run it. These are the prerequisites before you get started;

• Node installed on your local machine. You can use this command to check if it is installed;

node –version

If it is installed, you will get an output similar to this. 

If not, you can download Node. 

• Python from version 3.9+

python3 --version if you are using Ubuntu or python --version if running on Windows or macOS. 

If Python is already installed, you will have something similar to this; 

Node.js installs npm by default. You can now go ahead and install mermaid-js using this command; 

sudo npm install -g @mermaid-js/mermaid-cli

The final step is cloning the MetaGPT repository to your local machine. Use these commands; 

git clone https://github.com/geekan/metagpt

cd metagpt

python setup.py install

What are the use cases of MetaGPT?

Based on a research paper published by the creators of MetaGPT, you can use this framework for various tasks. These are some of the best;

Build games

You can create different games ranging from a snake, a flappy bird to a brick breaker game. Give MetaGPT a few prompts, and start playing your favorite game. 

Transcribe music

You can use MetaGPT to build an application that transcribes sheet music into a digital format. 

Custom press releases

Users can write a Python script that scraps a company’s data on social platforms and websites and creates custom press releases using the data collected. 

How MetaGPT’s adaptability can help enhance the multi-agent simulation complexity

Simulations are virtual experiments performed using models that imitate reality and are used for forecasting and analysis. MetaGPT is adaptable, making creating complex and realistic agent behaviors easy. This framework can learn from large amounts of data and generate behaviors likely to occur in the real world. 

MetaGPT is also programmed to analyze and understand the physical properties of the environment and how they affect the agents. This feature makes it easy for this framework to create realistic and hard-to-predict simulations. 

Comparing MetaGPT with Its Alternatives

MetaGPT is not the only framework utilizing various GPTs. Some of the popular alternatives are Python Read-Eval-Print Loop (REPL), LangChain, AutoGPT, and AgentVerse. We can compare MetaGPT with these frameworks on these fronts;

• Code generation: All the listed tools generate code. The only distinguishing feature of MetaGPT is that it offers a complete toolkit for project execution and management. 
• Code review: AgentVerse and MetaGPT are the only frameworks that have a code review feature. However, MetaGPT goes a step further and introduces the precompilation execution, making it easy to detect errors early. 
• API generation: Based on the features of the alternatives we have mentioned, MetaGPT is the only framework that offers API generation features. This feature makes it easy to prototype APIs during the software development process. 
• Collaboration: Role-based collaboration is available in AgentVerse and MetaGPT. This feature encourages multi-agent management and collaboration. 

Limitations of MetaGPT

Despite the endless possibilities that MetaGPT presents, it also lacks in these areas;

• Still in development: A close look at the project’s GitHub page shows it is still in active development. The project is yet to become perfect, but we can only wait and see how it will turn out in the future. 
• Not the perfect tool for complex projects: MetaGPT is a good framework for basic apps. However, you need a lot of human input when dealing with complex projects with a lot of data. 
• Limited to its training: Generative models can only produce data they have been trained on. The GPTs that MetaGPT is based on must be updated frequently to make it more accurate. 

FAQs

Conclusion

MetaGPT is designed to manage multi-agents through role definition, process standardization, and task decomposition. 

MetaGPT has many use cases. For instance, in software development, you can use MetaGPT to create software from scratch, improve existing software, generate user stories, and enhance collaboration. 

It is still too early to conclude whether MetaGPT is the best multi-agent framework. Even though the product is still in development, it has proved to be a good tool for the software development lifecycle. 

You may also read how to install Auto-GPT in minutes.