Paper Review | Understanding MetaGPT
Meta Programming For Multi-agent Collaborative Framework
MetaGPT is a quite long paper that exhaustively describes a quite straightforward approach to workflow management without providing much quantitative insights on the benefits of that approach.
Indeed, operating cognitive specialized agents in a software development context could progress into automating actual application development and might provide insights for further generalizations into real-world applications. Yet, I do believe that the approach has certain weaknesses.
Let me explain.
We already know that AutoGPT (and LangChain, AgentVerse) despite all its hype has not yet been able to live up to its expectations. I agree with the complaints by the authors of the paper that AutoGPT et al suffer from unproductive feedback loops and lack of guidance leading to no meaningful human/AI collaborations.
The paper argues that these existing agent systems tend to oversimplify complexities inherent to real-world applications because they are not embedded in advanced human management experience, and therefore are unable to solve larger and more complex real-world projects. And they are correct in their assumptions.
In general, the lack of workflow is exactly what most projects, especially the predecessor projects, are missing.
Many projects treat GPT as a brain that doesn’t need a body to function.
“MetaGPT is a meta-programming technology that utilizes SOPs to coordinate LLM-based multi-agent systems”
The authors further argue that the decomposition of human work experience as defined in corporate standard operating procedures (SOP) might be the way forward. And in general, I agree.
Standardizing SOPs should be capable of creating normative artifacts whose quality is high enough to work as connectors between the decomposed tasks. However, for most real-life applications, at least in 2023, that assumption is a bit naive but identifies a valuable friction point.
The project expresses its SOPs through the workflow within a shared environment. That makes sense.
But, it comes down to designing good agents within this shared "cognitive” environment. MetaGPT bases its agent definitions on roles. The design works abbreviated like this.