This arXiv preprint introduces ChatDev, a chat-powered software development framework in which specialized agents driven by large language models take part in the design, coding and testing phases of building software. The paper’s starting point is that software development is a complex task requiring cooperation among multiple members with diverse skills, and that earlier work applying deep learning to individual phases of a waterfall model gave each phase its own uniquely designed model — which the authors argue leads to technical inconsistencies across the various phases, and so to a fragmented and ineffective development process.
ChatDev’s answer is to route those phases through a single medium. Its agents are guided in what to communicate by a mechanism the paper calls a chat chain, and in how to communicate by one it calls communicative dehallucination. The agents actively contribute to design, coding and testing through unified language-based communication, with solutions derived from their multi-turn dialogues rather than from a separately designed model per phase.
The authors report an observation about which language suits which task: the agents’ utilization of natural language is advantageous for system design, while communicating in programming language proves helpful in debugging. They present the work as demonstrating how linguistic communication facilitates multi-agent collaboration, establishing language as a unifying bridge for autonomous task-solving among LLM agents.
The paper is filed under Software Engineering (cs.SE), Computation and Language (cs.CL) and Multiagent Systems (cs.MA). It was first submitted on 16 July 2023 and last revised on 5 June 2024 as version 5, carries the arXiv-issued DOI 10.48550/arXiv.2307.07924, and was accepted to ACL 2024. The code and data are stated to be available at https://github.com/OpenBMB/ChatDev.
Key Points
- The paper introduces ChatDev, a chat-powered software development framework whose specialized agents are driven by large language models and contribute to the design, coding and testing phases.
- Its stated motivation is that giving each waterfall-model phase its own uniquely designed deep learning model produces technical inconsistencies across phases, which the authors say results in a fragmented and ineffective development process.
- Agents are guided in what to communicate via a chat chain, and in how to communicate via communicative dehallucination — the paper’s two named mechanisms, introduced in the abstract by their role rather than described in detail there.
- The authors report that natural language is advantageous for system design, while communicating in programming language proves helpful in debugging — a split between the two media rather than a preference for one.
- The paper’s broader claim is that linguistic communication facilitates multi-agent collaboration, establishing language as a unifying bridge for autonomous task-solving among LLM agents; the abstract presents this as a paradigm-level conclusion rather than as a measured result.
Notes
The contribution is framed against earlier work that used deep learning to improve specific phases in a waterfall model — design, coding and testing taken one at a time. What ChatDev is offered as replacing is therefore not any single phase’s model but the arrangement of having one per phase, and the claim the authors rest on is about consistency across phases rather than about performance within one. The preprint went through five versions between July 2023 and June 2024.