====== ChatDev ======
**ChatDev** is an open-source multi-agent framework that automates software development by simulating a virtual company of LLM-powered agents collaborating through structured conversations(([[https://arxiv.org/abs/2307.07924|Qian et al. "Communicative Agents for Software Development" arXiv:2307.07924, 2023]])). Inspired by the waterfall model of software engineering, ChatDev assigns agents to roles such as CEO, CTO, programmer, and tester, enabling end-to-end software creation from a single natural language requirement. The project evolved from its initial 2023 release (v1.0) to the more flexible ChatDev 2.0 (DevAll) architecture.

  * **Website:** [[https://chatdev.ai|chatdev.ai]]
  * **[[github|GitHub]]:** [[https://github.com/OpenBMB/ChatDev|github.com/OpenBMB/ChatDev]]
  * **Paper:** [[https://arxiv.org/abs/2307.07924|arXiv:2307.07924]]
  * **Built on:** CAMEL framework

===== Agent Collaboration Model =====
Agents role-play as members of a software company hierarchy:

  * **CEO:** Sets vision, defines requirements from user input
  * **CTO / CPO:** Handles technical design, architectural decisions, and product planning
  * **Programmer:** Writes code based on specifications
  * **Tester:** Validates functionality through testing
  * **Reviewer:** Conducts code review and quality checks
  * **Art Designer:** Creates UI assets and visual elements

The core **ChatChain** mechanism breaks development into subtasks following a waterfall sequence. Each subtask involves multi-turn dialogues between agent pairs that work toward consensus. **Communicative dehallucination** prompts agents to seek clarification before responding, reducing errors like incomplete code or hallucinated functions.

===== Version 1.0 vs 2.0 =====
^ Aspect ^ Version 1.0 ^ Version 2.0 (DevAll) ^
| **Structure** | Fixed roles in virtual company | Decoupled, domain-agnostic workflows |
| **Interface** | Prompt-based orchestration | Visual drag-and-drop (Vue3 console) + Python SDK |
| **Workflow** | Linear waterfall ChatChain | DAG topologies with parallel tasks and dynamic routing |
| **Scope** | Software development only | Any complex multi-agent task |
| **Architecture** | Monolithic | Three-layer: Server (state), Runtime (execution), Workflow (logic) |

===== Results and Performance =====
  * Completes simple projects (e.g., Rock-Paper-Scissors game) in minutes
  * Cost-efficient: approximately $0.30 per simple project using GPT-3.5-turbo
  * Full SDLC automation: design, coding, testing, documentation from a single prompt
  * Communicative dehallucination significantly reduces [[llm_hallucination|LLM hallucination]]-related errors
  * Version 2.0 enables non-technical users via no-code tools while retaining developer control

===== Limitations =====
  * Quality dependent on underlying LLM capabilities
  * Complex projects may require human intervention for architectural decisions
  * LLM hallucinations remain a challenge despite mitigation strategies
  * Version 1.0's rigid waterfall structure limits flexibility (addressed in v2.0)

===== Related Pages =====
  * [[multi_agent_systems|Multi-Agent Systems]]
  * [[autogen|AutoGen]]
  * [[crewai|CrewAI]]
  * [[metagpt|MetaGPT]]
  * [[agent_protocol|Agent Protocol]]

===== See Also =====
  * [[opendev|OpenDev]]
  * [[how_to_build_a_chatbot|How to Build a Chatbot]]
  * [[flowise|Flowise]]
  * [[waydev|Waydev]]
  * [[chatgpt|ChatGPT]]

===== References =====