Visual Workflow Designer

ChatDev's flagship feature is its drag-and-drop workflow designer, built with Vue 3. You create agents as boxes on a canvas, connect them with directional edges, and configure each agent's role, LLM, system prompt, and tools through a visual inspector panel. The interface is reminiscent of Zapier or n8n, but purpose-built for AI agent orchestration. Each agent node can be assigned a different LLM backend — so you could have a research agent using Claude Opus for synthesis, a data extraction agent using GPT-4.1 for structured output, and a writing agent using Gemini 3 Flash for creative generation, all in the same workflow [3].

MacNet DAG Architecture

Introduced in June 2024, MacNet (Multi-Agent Collaboration Networks) is ChatDev's architectural foundation. It uses a Directed Acyclic Graph model where each agent is a node and edges represent one-way data flow. The DAG constraint prevents circular dependencies and infinite loops — a common failure mode in multi-agent systems. MacNet supports over 1000 agents in a single workflow without context window limits by splitting communication into modular sub-graphs. Each sub-graph operates within its own context window, with only summarized outputs passed between sub-graphs [4].

Iterative Experience Refinement (IER)

IER is ChatDev's most advanced feature and a genuine differentiator in the multi-agent space. When agents execute a workflow, they record what worked, what failed, and how they adapted. On subsequent runs, agents retrieve relevant past experiences and adjust their collaboration strategy. This means a research pipeline that failed to find relevant sources on the first run will try different search strategies on the second. Over time, agent teams develop "institutional memory" that improves reliability without human intervention. IER was introduced in a May 2024 paper and is deeply integrated into ChatDev 2.0's runtime [5].

Multi-Domain Template Library

ChatDev ships with a growing library of pre-built workflow templates organized by domain. The data visualization templates create 4-6 high-quality PNG charts from datasets. The 3D generation templates integrate with Blender via MCP to produce 3D assets. The research pipeline templates implement gather-analyze-write chains with citation management. The content creation templates handle the full workflow from topic research through draft generation to publication. Each template is a YAML file in the yaml_instance/ directory that users can customize through the UI or directly edit [1].

YAML-Based Workflow Configuration

Every ChatDev workflow is defined as a YAML file — a human-readable format that's more accessible than Python code while still being version-controllable. Environment variables can be injected via ${VAR} placeholders, keeping API keys out of workflow definitions. Users can switch between the visual designer and raw YAML editing, with changes synced in real-time. The validation command (make validate-yamls) checks all workflow files for correctness before deployment [1].

OpenClaw Integration

ChatDev 2.0 integrates natively with the OpenClaw agent ecosystem. You can install ChatDev skills via clawdhub install chatdev and then trigger workflows from within OpenClaw. Example prompts include "Create a ChatDev workflow to automatically collect trending information, generate a social media post, and publish it" — bridging the gap between local agent execution and multi-agent orchestration [1].