Chen Qian and ChatDev: A Software Company Made of Language-Model Agents
How ChatDev modeled software development as a sequence of conversations among specialized LLM agents for design, coding and testing.
In July 2023, Chen Qian and collaborators introduced ChatDev, a framework that treated software development as a multi-agent communication problem.[1] Specialized LLM agents took part in phases such as design, coding and testing, coordinating through structured conversations.
The metaphor was deliberately organizational: instead of one model trying to perform every task, the system resembled a small software company with distinct roles.
Natural language as the coordination layer
ChatDev’s key insight was that language could unify phases of software development that had traditionally required separate systems. Agents could discuss requirements, exchange implementation information and use programming language itself during debugging.[1]
This made communication both the strength and the risk of the architecture. Language is flexible, but it can propagate misunderstandings or hallucinations from one agent to another.
Software engineering becomes organizational design
ChatDev helped popularize a pattern that became common in later agentic coding systems: assign specialized responsibilities, define a communication workflow and let agents hand artifacts to one another. The code is only one output; plans, reviews and tests become intermediate products in an artificial organization.
That is a deeper change than simple code generation. The object being automated is the development process itself.
Why ChatDev matters historically
ChatDev was not the first multi-agent system and not the only 2023 LLM-agent framework. Its importance lies in applying the multi-agent idea directly to software engineering with an explicit role-based organizational metaphor.
For the history of coding, that makes ChatDev especially relevant: it foreshadowed a future in which programmers increasingly design the workflow by which agents produce software rather than writing every artifact themselves.
Works Cited
- 01
CodeHistory is a living archive. Citations document the evidence used for this edition; later evidence may refine the account.
Submit a research lead