# Open Dynamic Workflows: Open-Source Multi-Agent Workflow Orchestration Engine

> An MIT-licensed open-source project that implements the script-as-orchestrator multi-agent workflow pattern, supporting local execution, multi-model integration, and six workflow topology structures.

- 板块: [Openclaw Llm](https://www.zingnex.cn/en/forum/board/openclaw-llm)
- 发布时间: 2026-06-05T13:43:55.000Z
- 最近活动: 2026-06-05T14:53:06.121Z
- 热度: 160.8
- 关键词: multi-agent, workflow, orchestration, open-source, Claude Code, OpenCode, Codex, local-first, MIT license
- 页面链接: https://www.zingnex.cn/en/forum/thread/open-dynamic-workflows
- Canonical: https://www.zingnex.cn/forum/thread/open-dynamic-workflows
- Markdown 来源: floors_fallback

---

## Introduction / Main Floor: Open Dynamic Workflows: Open-Source Multi-Agent Workflow Orchestration Engine

An MIT-licensed open-source project that implements the script-as-orchestrator multi-agent workflow pattern, supporting local execution, multi-model integration, and six workflow topology structures.

## Original Author and Source

- **Original Author/Maintainer**: Suraj1235
- **Source Platform**: GitHub
- **Original Title**: open-dynamic-workflows
- **Original Link**: https://github.com/Suraj1235/open-dynamic-workflows
- **Release Date**: June 5, 2026
- **License**: MIT

---

## Project Overview

Open Dynamic Workflows is an open-source multi-agent workflow orchestration engine whose core innovation lies in adopting the **Script-as-Orchestrator** architectural pattern. This pattern originally appeared in Claude Code's dynamic workflows and ultracode features, and this project has open-sourced it under the MIT license, making it usable in various development environments such as OpenCode, OpenAI Codex, Google Antigravity, and VS Code.

Traditional multi-agent systems usually let a single large language model coordinate dozens of agents at the same time, which causes the model to consume a lot of context window on tracking the status of other agents. Open Dynamic Workflows' solution is to let the model **write the execution plan only once**—a normal JavaScript `execute()` function—then the local daemon is responsible for execution, and the model itself exits the execution loop.

---

## Script-as-Orchestrator Pattern

The core philosophy of this project can be summarized as: **The model is the author, the script is the orchestrator**. When a user describes a workflow task, the model generates a JavaScript script containing the `execute(context)` function, then hands it over to the local daemon for execution.

The daemon runs the script in a WASM-isolated QuickJS sandbox, where the only available scope is workflow primitives—including `agent`, `parallel`, `pipeline`, `verify`, `loop`, and `checkpoint`. Each `agent()` call becomes an HTTP request to the model provider, scheduled via a concurrent queue. The user's chat window will only see the final answer, not the intermediate processes.

## Local-First and Zero Telemetry

The project adheres to the Local-first principle:
- Supports self-hosted models (via Ollama)
- Zero telemetry; data does not leave the user's machine
- Supports multiple model providers (Anthropic, OpenAI, Ollama, etc.)
- Automatic model routing; intelligently selects providers based on model ID prefixes

---

## Six Workflow Topology Structures

Open Dynamic Workflows provides six workflow topologies, and the planner automatically selects the most suitable shape based on task characteristics instead of using the same pattern for all tasks:

## 1. MapReduce

Suitable for scenarios where the same check is performed on a large number of items, such as auditing 500 files. The workflow is: Split → Parallel Map → Reduce Summary.

## 2. Pipeline

Suitable for tasks that require sequential execution of multiple stages, such as Migration → Testing → Fixing. Each item can flow independently without waiting for other items to complete the previous stage.