ChatDev: Communicative Agents for Software Development

“ChatDev: Communicative Agents for Software Development” was posted to arXiv on July 16, 2023 by a team led by Chen Qian, with authors from Tsinghua University and collaborators, and was accepted at ACL 2024. It frames software development as a sequence of conversations among specialized language model agents that take on roles like CEO, programmer, and reviewer and move a project through design, coding, and testing phases.

The framework’s organizing device is the chat chain, which decomposes each development phase into a series of small, well-defined subtasks and routes each one through a focused dialogue between two roles. ChatDev also introduces communicative dehallucination, a pattern in which an agent requests more detail before committing to a response, to cut the errors that accumulate when agents act on under-specified instructions. The authors observed that natural language works well for the high-level design discussion while communicating in programming language is what helps during debugging, and they offered the system as evidence that language can serve as a unifying bridge for autonomous task-solving among agents.

ChatDev and MetaGPT are the two most-cited demonstrations of the “software company as a team of agents” idea from 2023, and both fed the broader push - later carried into benchmarks like SWE-bench and agents like SWE-agent - to measure whether agent teams can actually ship working code.

Sources

Last verified June 7, 2026