Intel / repo

ChatDev 2.0

Source: arxiv.org
Author: Chen Qian et al.
Date: Indexed Apr 14, 2026

Photo by Rafael Rex Felisilda on Unsplash

Context

This item links to the original ChatDev research paper, ChatDev: Communicative Agents for Software Development, first posted in July 2023 and accepted to ACL 2024. The paper's question is whether a team of language-model agents, each given a job role, can build working software by talking to one another instead of relying on a separate specialised model for each stage of development.

ChatDev organises that conversation with what the authors call a chat chain. Development is split into design, coding and testing phases, which are broken into smaller subtasks. Each subtask pairs two agents with roles such as CEO, CTO, programmer, reviewer or tester: an instructor gives directions and an assistant proposes solutions, and the pair talks over several turns until they agree. Only the agreed solution, not the whole dialogue, is carried into the next phase, which keeps later agents from drowning in earlier chatter.

The second idea, communicative dehallucination, targets code that is incomplete, will not run, or drifts from the requirements. Before answering an instruction, the assistant first asks the instructor for specifics, such as the exact name of a dependency, and only then commits to a response. In the paper's own comparison, ChatDev scored higher than GPT-Engineer and MetaGPT on completeness, executability, consistency and an overall quality score, though it took far longer and used more than GPT-Engineer.

The item's title refers to ChatDev 2.0, but the linked document is this earlier paper. According to the ChatDev repository, ChatDev 2.0 was released in January 2026 as a broader zero-code platform for configuring workflows, and the role-playing software company described in the paper is now maintained as ChatDev 1.0, which the project calls the foundation of its approach.

Terms in this piece · Glossary

token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
multi-agent — Using several AI agents on one problem — splitting work in parallel, checking each other, or filling different roles like planner and reviewer.

Key quotes

“In this paper, we introduce ChatDev, a chat-powered software development framework in which specialized agents driven by large language models (LLMs) are guided in what to communicate (via chat chain) and how to communicate (via communicative dehallucination).”

“We found their utilization of natural language is advantageous for system design, and communicating in programming language proves helpful in debugging.”

“This paradigm demonstrates how linguistic communication facilitates multi-agent collaboration, establishing language as a unifying bridge for autonomous task-solving among LLM agents.”

Read the source arxiv.org