ChatDev 2.0 DevAll: Open-Source Multi-Agent Platform

ChatDev 2.0, also called DevAll, is a meaningful shift in how open-source AI software projects are packaged. The original ChatDev presented the metaphor of a virtual software company: specialized agents such as a CEO, CTO, programmer, and tester collaborated through structured conversations. The new release keeps the multi-agent idea but turns it into a broader, configurable orchestration platform that can be used without writing an agent framework from scratch.

What ChatDev 2.0 actually changes

The project describes DevAll as a zero-code multi-agent platform for “developing everything.” Instead of limiting the system to software generation, users define agents, workflows, and tasks through configuration and then execute scenarios such as data visualization, 3D generation, and deep research. That is a useful distinction. Many agent demos are really single prompts wrapped in a user interface; DevAll is focused on making the workflow itself a first-class artifact.

The repository now includes a backend, a Vue 3 frontend, workflow definitions, a schema registry, MCP examples, runtime components, tests, and Docker-related files. The project’s README says ChatDev 2.0 was officially released on January 7, 2026, while the older software-company implementation remains available as the maintained ChatDev 1.0 legacy branch.

For developers, this separation matters. It lets the newer platform evolve toward configurable orchestration without erasing the older research and teaching example. Teams can study the original role-based collaboration model while using DevAll as a more general runtime.

The architecture: configuration before code

The quickest way to understand DevAll is to think in terms of a control plane for agent teams. A user provides model credentials and a base URL, defines workflows in YAML, and runs the backend and web console. The project supports Python 3.12 or newer, Node.js 18 or newer, and uses uv for Python dependency management. Its documented development path starts the backend on port 6400 and the Vite/Vue frontend on port 5173.

The project also provides commands for synchronizing YAML workflows to the frontend and validating those workflows. This is more than a convenience feature. Validation creates a boundary between an agent design and an execution attempt. If a workflow is invalid before it reaches the runtime, the failure is easier to diagnose than a silent breakdown in the middle of a long autonomous task.

The repository’s OpenClaw integration is another notable detail. ChatDev says OpenClaw can invoke existing agent teams or dynamically create new teams. Its example use case is an automated information-collection workflow that gathers trending information, generates a social post, and publishes it. That pattern maps directly to real content and research pipelines, where retrieval, synthesis, review, and delivery should be separate steps rather than one opaque model call.

Why multi-agent orchestration still needs discipline

Adding agents does not automatically improve a system. Every additional role can introduce latency, duplicated context, conflicting instructions, and extra opportunities for an incorrect action. The value of DevAll is therefore not simply that it supports many agents. Its value is that it makes the collaboration topology configurable and inspectable.

A production workflow should assign each agent a narrow responsibility. For example:

1. A researcher gathers source material and records URLs.
2. A verifier checks dates, claims, and licensing language.
3. A writer produces a structured draft from the verified notes.
4. A reviewer tests the draft against editorial and SEO requirements.
5. A publisher performs the final external action only after approval.

This arrangement also creates useful failure boundaries. If the verifier cannot confirm a claim, the writer should receive a blocked or qualified result instead of a confident invention. If the publisher fails, the research and draft artifacts should remain available for retry.

ChatDev’s research lineage

DevAll is built on a longer line of work around communication between language-model agents. The original ChatDev used role-based seminars to automate parts of the software development lifecycle, including design, coding, testing, and documentation. The repository also points to later research on evolving orchestration, a puppeteer-style paradigm, and multi-agent collaboration networks.

That history explains why the project is interesting beyond its interface. The central question is not merely “how do I call several models?” It is “how should a system decide which agent acts next, what context it receives, and when collaboration should stop?” Directed workflows, role specialization, and reusable configurations are practical answers to that question, even when the underlying models change.

A practical evaluation plan

If you want to test ChatDev 2.0, start with a workflow that has a measurable output and a low-risk sandbox. A good first experiment is a small research-and-report task:

• Give one agent a fixed set of source URLs.
• Ask a second agent to extract factual claims with citations.
• Have a third agent produce a report in a strict schema.
• Run automated checks for missing sources, unsupported numbers, and malformed output.
• Compare the multi-agent result with a carefully prompted single-agent baseline.

Measure total runtime, model calls, token usage, correction rate, and the percentage of claims that survive verification. The baseline is essential: if three agents do not improve accuracy or maintainability, the extra orchestration is not justified.

For software generation, use a repository with tests and ask the workflow to produce a patch rather than directly deploy. Require the final agent to run the test suite and return the diff, logs, and unresolved risks. Human approval should remain between generated changes and production systems.

The bigger developer takeaway

ChatDev 2.0 reflects a broader trend in AI development: the competitive layer is moving from prompt templates to durable execution systems. Persistent state, workflow schemas, tool permissions, replayable runs, validation, and approval gates increasingly determine whether an AI system is dependable.

DevAll will not remove the hard parts of agent engineering. Teams still need to choose models, control costs, protect credentials, sandbox generated code, and design recovery paths. But by making agent teams and workflows configurable, the project gives developers a useful laboratory for those problems.

The most promising use of ChatDev 2.0 is not asking a crowd of agents to improvise. It is expressing a clear process—research, transform, verify, test, and approve—as a reusable workflow that can be inspected and improved over time. That is the direction AI developer tooling is taking, and DevAll is a substantial open-source project to watch as that direction matures.

Related reading

• Microsoft Agent Framework Harness Guide
• Nanobot: Open-Source Personal AI Agent Runtime
• OpenAI Agents SDK 0.20: MCP v2 Migration Guide