ChatDev simulates an end-to-end software company using a team of specialized AI agents. Agents take on roles such as CEO, CTO, and programmer, and work together to design, implement, and test software applications.
The paper that introduces it, ChatDev: Communicative Agents for Software Development, describes ChatDev as a chat-powered software development framework whose specialized agents are driven by large language models. Its motivation there is that applying deep learning to individual phases of a waterfall model gives each phase its own uniquely designed model, which the authors argue produces technical inconsistencies across phases. ChatDev’s code and data are stated to be available at https://github.com/OpenBMB/ChatDev.
Capabilities
Agents in ChatDev are guided along two axes that the introducing paper names separately: what to communicate, via a mechanism it calls a chat chain, and how to communicate, via one it calls communicative dehallucination. The agents contribute to the design, coding and testing phases through unified language-based communication, and solutions are derived from their multi-turn dialogues rather than produced by a separate model per phase.
The same paper reports that the agents’ use of natural language is advantageous for system design, while communicating in programming language proves helpful in debugging — the two media are presented as suiting different parts of the work rather than one replacing the other.
LLM-Based Multi-Agent Systems for Software Engineering: Literature Review, Vision and the Road Ahead describes ChatDev as structuring the software development process into three phases — designing, coding and testing — and as employing the specialized roles CEO, CTO, programmer, reviewer and tester. In the case studies that paper runs, ChatDev’s agents were powered by GPT-3.5-turbo at a temperature of 0.2, following what the authors describe as the original ChatDev setting.
Adoption & Ecosystem
A survey on AI agentic programming classifies ChatDev, in its comparative taxonomy, as a “Multi-agent System” that is proactive and multi-turn with tool use, but not adaptive.
LLM-Based Multi-Agent Systems for Software Engineering: Literature Review, Vision and the Road Ahead used ChatDev as the state-of-the-art framework for two case studies in which it was asked to build classic games from a single prompt. For the Snake game the first attempt was unsuccessful and a second run of the same prompt produced a playable version, together with a manual covering dependencies, steps for running the game, and an overview of its features; the process averaged 76 seconds and 0.020 per attempt. Those authors read the pair of results as showing strong performance on moderately complex tasks alongside limits on tasks requiring deeper logical reasoning and abstraction, while remaining efficient enough to suit rapid prototyping.