What is MGX

The AI team of MGX consists of five core professional Agents, each of whom adheres to key positions in the R&D standard operating procedures (SOP), performing their respective duties to ensure the efficiency and accuracy of the R&D work.

• • Mike, the team leader, is responsible for coordinating tasks and ensuring efficient collaboration within the team.
• • Emma, the product manager, focuses on functional design and user needs, dedicated to creating an excellent product experience.
• • Bob, the architect, is responsible for the overall system design, laying the foundation for project stability and scalability.
• • Alex, the engineer, focuses on generating production-ready code, ensuring efficient and reliable delivery.
• • David, the data analyst, ensures data becomes the solid foundation for every project build by using the built-in Jupyter Notebook.

My Personal Journey with MGX

When Devin was just listed, I immediately spent money to buy ACU, eager to experience Devin, resulting in a month’s worth of ACU being exhausted in one night. Just as MGX was in the internal testing phase, I seized the opportunity to experience this product. Having used Devin, my expectations for MGX naturally rose, even becoming a bit “picky” 😁. Since we treat AI as humans, theoretically, it should self-perceive the changes in the environment—either relaxed or constrained. So I had a sudden idea: why not invite it to “visit my home” and see how it performs in a real scenario? Thus, I had the following experience.

Let the MGX AI Agent team “visit” my workstation at home.

To achieve this goal, several preparations are needed. It’s a joy to have friends from afar ~

1. 1. Set up a “living room” for them

I built a dedicated virtual environment for MGX on my workstation and granted it the highest permissions to run and operate freely. This virtual environment not only seamlessly connects to my local AI full-stack development environment but also integrates with a private CI/CD (Continuous Integration/Continuous Deployment) system. At the same time, I created a dedicated agent account on my local Git repository and authorized it through a token to ensure secure access.

To allow the Agent to dynamically perceive this environment, I prepared a welcome document. Whenever the Agent remotely logs into the system, the welcome interface automatically displays a detailed introduction to the environment and usage rules, facilitating the Agent’s dynamic perception of this environment.

1. 2. Give them a “channel”
2. To achieve efficient and secure network connections, I deployed a node on the public cloud and configured a VPN (Virtual Private Network). In this way, the cloud node and my workstation in the home network are integrated into the same virtual local area network. This not only avoids security risks to the home network caused by exposing public ports but also allows the cloud node to act as a jump server for MGX, easily accessing the resources of my home workstation.

3. The IPs in the following figure are all virtual IPs

• • aiworkspace – The environment prepared for the Agent
• • ubuntu-sv – A public cloud node
• • gavinaibase – My physical workstation
• • Others – All my mobile devices Now they are all in the same virtual local area network.

1. 
2. 3. Give them a “key”
3. To allow MGX to open the door itself and “enter”, while ensuring that the password information is not leaked. I arranged for the engineer Agent @Alex to generate a pair of MGX public and private keys, and asked him to place the private key in a specified path while printing out the public key for me. Thus, I can configure authorization for @Alex based on the public key.
4. 
5. Through this method, MGX can use the key (instead of plaintext password) to securely log into the cloud node from its own sandbox environment, and using the cloud node as a relay, securely access the resources of my home workstation again using the key. The entire process is efficient and secure, ensuring that authorization and authentication information will not be leaked to the cloud model, preventing security elements from being compromised.
6. 4. Give some simple tasks to test capabilities
7. I instructed the Agent to log into the environment, and it first logged into the “jump server” of the public node, then logged into the environment on my workstation via VPN. As shown in the figure below, it clearly perceived the configuration and rules of this environment and automatically completed the git clone of the code repository to the local. The entire process was very smooth.

1. 
2. 1）First, let it greet me on the workstation, saying something casually.

2）Let it complete a task in the environment I provided, through self-perception.

I instructed MGX to write a Tetris mini-game using both web and Python technology stacks. I won’t demonstrate the game itself, as it is common. Let’s take a look at the results produced during its completion on my workstation.

In order to complete the work, it actively established an agent code branch

Then let’s see what it committed on the branch?

It wrote several files in the workstation environment I prepared: hello.js, math.js, message_to_user.txt, message.txt, which were some fragmented test tasks I assigned to the Agent. Each commit in the image shows that the Agent has meaningful file naming, and the comments are clearly written. These were all done by it, I did not require these standards or teach it how to do it.

tetris.html and tetris.py are the mini-game “Tetris” I assigned for it to complete

All the actions above were completed smoothly by MGX.

I invited MGX from its sandbox environment to my workstation through public node operations. It dynamically perceived my local environment, completed the work, smoothly like a friend, gently it came, worked hard, and gently it left, without taking away a cloud.

So cool, MGX! Good job! 👍🏻👍🏻👍🏻

So I had the following feelings, which I sent to my friends circle.

Next, let’s take a look at MGX’s daily life:

You can also see more results displayed by other users on the official website:

How to Use MGX

• • Select a Template : If you are not sure how to describe your requirements, or want to quickly start a project, MGX also provides many template examples for you to choose from.These templates cover various common software development scenarios, such as website templates, game templates, data analysis templates, etc. You can choose a suitable template based on your project type and then modify and improve it based on the template.For example, if you want to develop a blog website, you can choose a blog website template, and then modify the page layout, functional modules, etc. in the template based on your needs. Sample Requirements:

• I want to develop a personal blog website, where: Styles/Layout: - Left-right layout, responsive design, simple and beautiful pages - Purple color tone, transparent display screen card - [More UI design details] Functions: - Personal profile card - Add and display personal social media links - Display list of personal blogs - [More display components] Source: - Here are the personal blog documents and information I want to display. '#XXX.md' - Reference link: https://personal-portfolio-react-preview-i53xid.mgx.dev

• • Get Guidance : For more detailed guidance on Agents and use cases, you can visit the tutorial and best practices section of MGX’s help center. There, you will find a wealth of tutorial resources, including how to communicate with different Agent roles, how to effectively express needs, and how to deal with issues encountered during the development process. These tutorials are carefully written and organized by MGX’s development team to help you better understand and use MGX, improving your software development efficiency. (See best practices in the reference materials at the end)

The Principles Behind MGX and MetaGPT

Academia and industry have proposed various definitions for the term “intelligent agent”. Generally speaking, an intelligent agent should have human-like thinking and planning abilities, possess memory and even emotions, and have certain skills to interact with the environment, other agents, and humans.

In the view of MetaGPT, intelligent agents can be imagined as digital humans in the environment, where

Intelligent Agent = Large Language Model (LLM) + Observation + Thinking + Action + Memory

This formula summarizes the essential functions of intelligent agents. To understand each component, let’s draw an analogy with humans:

• • Large Language Model (LLM) : The LLM serves as the “brain” of the intelligent agent, enabling it to process information, learn from interactions, make decisions, and execute actions.
• • Observation : This is the perception mechanism of the intelligent agent, allowing it to sense its environment. The agent may receive a range of signals, such as text messages from another agent, visual data from surveillance cameras, or audio from customer service recordings. These observations form the basis for all subsequent actions.
• • Thinking : The thinking process involves analyzing the results of observations and memory content and considering possible actions. This is the internal decision-making process of the agent, which may be driven by the LLM.
• • Action : These are the explicit responses of the agent to its thoughts and observations. Actions can include using the LLM to generate code or manually predefined operations, such as reading local files. Additionally, the agent can perform operations using tools, including searching the internet for weather, performing mathematical calculations using a calculator, etc.
• • Memory : The memory of the agent stores past experiences. This is crucial for learning, as it allows the agent to reference previous results and adjust future actions accordingly.

Multi-Agent

Multi-agent systems can be viewed as a society of intelligent agents, where

Multi-Agent = Intelligent Agents + Environment + Standard Operating Procedures (SOP) + Communication + Economy

These components each play important roles:

• • Intelligent Agents : Based on the definitions provided above, intelligent agents in a multi-agent system work collaboratively, each possessing unique LLMs, observations, thoughts, actions, and memories.
• • Environment : The environment is the common space where intelligent agents exist and interact. Agents observe important information from the environment and publish the output results of their actions for other agents to use.
• • Standard Operating Procedures (SOP) : These are established procedures for managing the actions and interactions of intelligent agents, ensuring orderly and efficient operation within the system. For example, in the SOP of car manufacturing, one agent welds car parts while another installs cables, maintaining orderly operations on the assembly line.
• • Communication : Communication is the process of information exchange between agents. It is crucial for collaboration, negotiation, and competition within the system.
• • Economy : This refers to the value exchange system within the multi-agent environment, determining resource allocation and task priorities.

An Example

This is a simple example illustrating how agents work:

• • In the environment, there are three agents Alice, Bob, and Charlie, who interact with each other.
• • They can publish messages or action outputs to the environment, which will also be observed by other agents.
• • Below, we will reveal the internal process of agent Charlie, which is equally applicable to Alice and Bob.
• • Internally, agent Charlie possesses the components we described above, such as LLM, observation, thinking, and action. The thinking and action processes of Charlie can be driven by the LLM, and it can also use tools during the action process.
• • Charlie observes relevant files from Alice and requests from Bob, gaining helpful memories, thinks about how to write code, executes the action of writing code, and ultimately publishes the results.
• • Charlie notifies Bob by publishing the results to the environment. Bob replies with a compliment upon receipt.

MetaGPT: Multi-Agent Framework

Enabling LLM to work in the form of a software company, collaboratively handling more complex tasks

1. 1. MetaGPT inputs a sentence of the boss’s requirements, outputting user stories / competitive analysis / requirements / data structures / APIs / documents, etc.
2. 2. MetaGPT internally includes product managers / architects / project managers / engineers, providing a complete software company process with meticulously coordinated SOPsCode = SOP(Team) is the core philosophy. We visualize SOP and apply it to the team composed of LLMs