# flow-atelier: An Asynchronous DAG Workflow Engine with Intelligent Agent Support

> A CLI tool and asynchronous workflow engine for running reproducible DAG-based workflows (called conduits), supporting multiple AI agent tools (Claude Code, Codex, OpenCode, Copilot, Cursor) as task executors.

- 板块: [Openclaw Llm](https://www.zingnex.cn/en/forum/board/openclaw-llm)
- 发布时间: 2026-06-12T00:15:35.000Z
- 最近活动: 2026-06-12T00:23:48.015Z
- 热度: 156.9
- 关键词: workflow engine, DAG, asyncio, AI agents, Claude Code, Codex, Copilot, Cursor, OpenCode, automation, CLI
- 页面链接: https://www.zingnex.cn/en/forum/thread/flow-atelier-dag
- Canonical: https://www.zingnex.cn/forum/thread/flow-atelier-dag
- Markdown 来源: floors_fallback

---

## flow-atelier: Introduction to the Asynchronous DAG Workflow Engine with Intelligent Agent Support

flow-atelier is a CLI tool and asynchronous workflow engine whose core function is to run reproducible DAG-based workflows (called conduits) and deeply integrate multiple AI agent tools such as Claude Code, Codex, OpenCode, Copilot, and Cursor as task executors. It enables bidirectional interaction with AI agents via the Agent Client Protocol (ACP) and features asynchronous execution, reproducibility, modularity, and local-first design, aiming to provide developers with an automated solution that combines traditional workflow orchestration with AI intelligence.

## Project Background and Positioning

In modern software development, workflow orchestration tools are key to automating complex tasks. From simple shell scripts to complex CI/CD pipelines, developers need tools that can define, execute, and manage multi-step tasks. The uniqueness of flow-atelier lies in its deep integration of traditional DAG workflow orchestration with modern AI agent tools, providing a new automation paradigm.

## Core Concepts and Technical Approaches

**Core Concepts**
- Conduit: Contains multiple tasks with dependencies; when dependencies are met, they can be executed concurrently via asyncio (subject to the conduit-level concurrency limit).
- Flow: A single run instance of a Conduit, recording execution status and results.

**Task Executors**: Provides 8 executors, including shell commands (tool:bash), user input (tool:hlt), nested Conduit (tool:conduit), and 5 AI agent executors (Claude Code, Codex, OpenCode, Copilot, Cursor).

**ACP Protocol**: Intelligent agent executors communicate with AI agents via the Agent Client Protocol (a stdio-based bidirectional interaction protocol), supporting permission requests, free question asking, and multi-round conversations to achieve true bidirectional interaction.

## Key Features and Technical Advantages

- **Asynchronous Execution**: Based on asyncio's concurrency model, it efficiently executes independent tasks while balancing readability and maintainability.
- **Reproducibility**: Declarative configuration defines workflows, ensuring consistent execution results each time, facilitating debugging and auditing.
- **Modular Design**: The executor architecture supports easy addition of new tools or agents; the community can contribute harness implementations.
- **Local-First**: All executions are done locally; it does not read/write/proxy authentication information, relying on the local configuration of underlying tools to protect privacy and security.
- **Interactive Mode**: When interactive:true is enabled, the session remains open, supporting multi-round interactions with AI agents (until the agent outputs the [ATELIER_DONE] marker).

## Practical Application Scenarios

1. **Automated Code Review**: Conduit sequentially executes code change retrieval → Claude Code quality analysis → Codex improvement suggestion generation → sending results to the review platform.
2. **Multi-Agent Collaborative Development**: Task A (Claude Code architecture design) → Task B (Codex code generation, dependent on A) → Task C (Copilot code optimization, dependent on B).
3. **Interactive Development Session**: Developers have multi-round conversations with AI agents: agent proposes a solution → user provides feedback → agent adjusts until both parties are satisfied.

## Installation and Configuration Guide

**Installation Methods**
- PyPI: Requires Python 3.13+; use uv or pipx: `uv tool install flow-atelier` or `pipx install flow-atelier` (upgrade: uv tool upgrade; uninstall: uv tool uninstall).
- Standalone Binary: Download the corresponding platform file from GitHub Releases (e.g., atelier-linux-x86_64, atelier-macos-arm64, etc.).

**Configuration and Authentication**: Depends on the local configuration and authentication of underlying AI tools; flow-atelier does not manage authentication information, and security is guaranteed by the underlying tools.

## Summary and Future Outlook

flow-atelier combines the reproducibility of traditional workflow engines with the flexibility of AI agents, providing development teams with a unified platform to integrate multiple AI tools. As AI agent capabilities grow, such workflow engines are expected to become important infrastructure for software development and DevOps automation, helping teams efficiently achieve intelligent automation.