MetaGPT turns AI agents into an assembly line, each one producing a fixed document before passing the work to the next.

This explores how MetaGPT and ChatDev assign agents fixed roles and lock them into a preset way of working together. One caveat first: the collection covers MetaGPT directly but has no note on ChatDev. Anything said about ChatDev here is inferred from the contrast MetaGPT draws, not taken from the corpus.

MetaGPT's core move is to copy human Standard Operating Procedures (SOPs) into the agent architecture Does structured artifact sharing outperform conversational coordination?. Each role is defined less by its persona than by the document it has to produce. The product manager writes a requirements doc, the architect writes a system design, and the engineer writes code against that design. The collaboration algorithm is the order those documents get handed off in. Agents don't chat freely. They publish structured artifacts to a shared environment and pull only what their role needs. The finding is that this beats conversational coordination, because free-form chat between agents builds up noise and drift that a fixed document schema filters out. ChatDev is usually described as the conversational counterpart: role pairs talk their way through each development phase. MetaGPT's result works as an argument against that approach.

The surprising lesson is that the role label itself does less work than you'd expect. Role-play prompting alone measurably changes model behavior: casting ChatGPT in a two-stage role-play dialogue raised zero-shot reasoning accuracy by 10 to 60 points across benchmarks Does role-play dialogue boost reasoning accuracy in ChatGPT?. So roles do help. But MetaGPT suggests the bigger gain comes from what the role is required to output and when. A role tied to a fixed artifact and a fixed point in the pipeline is a contract, not a character.

Other parts of the collection show the trade-off this rigidity creates. An executive LLM can steer a collaborative conversation toward specific goals while still letting it feel natural Can AI teammates assess collaboration without losing naturalness?. That's a middle path where structure lives in a supervising layer instead of in hard-coded handoffs. Realtime-Venus keeps one ordered shared record of delegated tasks and ongoing dialogue Can frontends handle delegation while staying conversationally engaged?. It's the same instinct as MetaGPT's shared environment, applied to human-facing agents. Generated UIs show the general cost of structure: clearer outputs, but more rigidity and harder changes mid-workflow Do generated analysis UIs really work better than chat?. Fixed role pipelines likely face the same tension: reliable on tasks that fit the SOP, brittle on tasks that don't.

What you might not have expected to learn: the main design choice in these frameworks isn't which roles exist. It's whether agents coordinate by talking or by passing structured documents, and the evidence here favors documents.