ChatDev

Enables declarative workflow definition through YAML configuration files stored in yaml_instance/ directory, eliminating code-based agent choreography. The runtime dynamically parses YAML schemas to instantiate agent nodes, configure tool bindings, and manage context flow between agents without requiring Python/JavaScript programming. Uses a configuration-driven architecture where workflow topology, agent roles, and data dependencies are expressed as structured YAML, then executed by a domain-agnostic orchestration engine that interprets node definitions and manages inter-agent communication.

Provides a browser-based Web Console (port 5173) with interactive workflow canvas enabling visual agent node composition, connection, and parameter configuration through drag-and-drop UI. The frontend layer communicates with the backend API layer to persist workflow definitions, validate node connections, and preview execution flow. Users visually design agent topologies by placing nodes representing agents/tools, connecting them to define data flow, and configuring node parameters through form-based UI without touching YAML directly.

Provides an abstraction layer for memory/knowledge storage enabling pluggable backends (database, vector store, file system) without modifying workflow definitions. Agents can store and retrieve information through a unified memory interface, with the actual persistence mechanism configured at runtime. Supports both short-term context memory (within workflow execution) and long-term knowledge storage (across executions), enabling agents to build cumulative knowledge and reference historical information.

Provides specialized workflow templates for software development, data visualization, 3D generation, game development, and research domains, each with pre-configured tool bindings, agent roles, and orchestration patterns. Templates encode domain expertise through predefined agent responsibilities (e.g., architect, developer, reviewer for software dev) and tool selections (e.g., code generation, testing, documentation tools). Users instantiate templates through YAML configuration, customizing domain-specific parameters while reusing proven orchestration patterns.

Enables batch processing of multiple workflow instances with parameter variation through Python SDK, executing workflows across datasets or parameter ranges and aggregating results. The batch system manages workflow instance lifecycle (creation, execution, result collection), supports parallel execution with configurable concurrency, and provides structured result aggregation enabling analysis across batch runs. Supports parameter sweeps, dataset iteration, and conditional batch execution based on previous results.

Provides an interactive tutorial interface within the Web Console enabling users to learn ChatDev through guided workflows, interactive examples, and step-by-step agent execution visualization. The tutorial system walks users through workflow concepts (agents, tools, context flow) with executable examples, showing how agents collaborate and how data flows through workflows. Users can pause execution, inspect agent state, and modify workflows in real-time to understand ChatDev mechanics.

Provides a monitoring dashboard within the Web Console displaying real-time workflow execution status, agent progress, resource utilization, and execution metrics. The dashboard shows active workflows, completed executions with results, and historical execution trends. Users can launch new workflow instances, monitor execution progress, view agent logs, and retrieve results through a unified interface. Supports filtering, searching, and exporting execution history for analysis.

Provides pre-built workflow templates for five distinct domains: software development, data visualization, 3D generation, game development, and deep research/video generation. Each domain template encodes domain-specific agent roles, tool bindings, and orchestration patterns that can be instantiated and customized through YAML configuration. The runtime loads domain-specific tools and LLM provider configurations based on the selected template, enabling the same orchestration engine to execute fundamentally different workflows without domain-specific code branches.

Exposes a Python SDK enabling programmatic workflow instantiation, execution, and batch processing for CI/CD integration and production automation. The SDK provides APIs to load YAML workflows, inject runtime parameters, execute workflows synchronously or asynchronously, and retrieve structured execution results. Supports batch processing of multiple workflow instances with parameter variation, enabling automation of repetitive tasks like code generation, data processing, or content creation at scale without manual intervention.

Enables developers to define custom agent nodes, tool providers, and memory backends through a plugin architecture without modifying core runtime. Custom nodes are registered in the extension system and referenced in YAML workflows, allowing teams to extend ChatDev with domain-specific agents, proprietary tools, or specialized LLM providers. The runtime dynamically loads custom node implementations at workflow instantiation time, enabling composition of built-in and custom agents in the same workflow.

Implements a context management system where agent outputs are automatically captured and passed as inputs to downstream agents based on workflow topology. The runtime maintains a context dictionary that flows through the workflow, with each agent node reading from context, executing its task, and writing results back to context. Supports variable injection and environment parameter passing, enabling agents to access workflow-level configuration, previous agent outputs, and external data without explicit parameter passing in YAML.

Implements a learning mechanism where agents capture execution experiences (successful patterns, failure modes, optimization insights) and refine their behavior across multiple workflow runs. The ECL system maintains an experience repository that agents query during execution to inform decision-making, and updates this repository based on execution outcomes. Enables agents to improve their orchestration strategies and tool selection over time without explicit retraining, using reinforcement learning principles to optimize workflow execution.

Implements a multi-agent architecture where agents are assigned specific roles (e.g., analyst, developer, reviewer) and coordinate through a structured communication protocol. The MacNet architecture defines role-specific responsibilities, communication patterns, and decision-making authorities, enabling complex multi-agent workflows where agents collaborate based on their assigned roles. Agents communicate through a message-passing system that respects role hierarchies and communication constraints, enabling emergent coordination without explicit choreography.

Implements an iterative refinement loop where workflows are executed, outcomes are analyzed, and workflow definitions are automatically adjusted to improve performance. The IER system compares actual execution results against expected outcomes, identifies bottlenecks or failure points, and suggests or applies YAML modifications to optimize the workflow. Enables workflows to self-optimize over multiple iterations without manual intervention, using feedback loops to discover better agent orderings, tool selections, or parameter configurations.

Implements a reinforcement learning-based orchestration system (Puppeteer) that dynamically schedules agent execution and resource allocation based on workflow state and execution history. The Puppeteer system learns optimal scheduling policies through RL training, enabling dynamic agent ordering, parallel execution decisions, and resource prioritization without explicit YAML choreography. Agents are scheduled based on learned policies that maximize workflow efficiency metrics (latency, cost, quality), adapting to changing conditions and agent capabilities.