Open Multi-Agent — TypeScript Multi-Agent Orchestration Framework

Introduction

Open Multi-Agent is a TypeScript framework for orchestrating multiple LLM agents to accomplish complex goals. You describe a high-level objective, and a coordinator agent decomposes it into a directed acyclic graph (DAG) of subtasks, assigning each to a specialized agent that can use any LLM provider.

What Open Multi-Agent Does

• Decomposes complex goals into task DAGs automatically via a coordinator agent
• Assigns subtasks to specialized agents with different roles and capabilities
• Supports any LLM provider: Claude, ChatGPT, Gemini, DeepSeek, or local models
• Manages inter-agent communication and result aggregation
• Provides built-in tool use so agents can call external APIs and functions

Architecture Overview

The framework centers on a coordinator that receives a goal and plans a task graph. Each node in the graph is a subtask assigned to a worker agent with a specific role (researcher, coder, reviewer, etc.). The coordinator handles dependency resolution, parallel execution of independent subtasks, and result synthesis. Agents communicate through a shared context, and each can use tools registered at the framework level.

Self-Hosting & Configuration

• Install via npm and import into any TypeScript or JavaScript project
• Configure LLM providers with API keys through environment variables
• Define custom agent roles and specializations via the agent configuration API
• Register tools (functions, APIs) that agents can invoke during task execution
• Works with Node.js 18+ and modern JavaScript runtimes

Key Features

• Model-agnostic: mix providers within a single orchestration run
• Automatic task decomposition eliminates manual workflow definition
• DAG-based execution maximizes parallelism across independent subtasks
• Type-safe TypeScript API with full IntelliSense support
• Extensible tool system lets agents interact with external services

Comparison with Similar Tools

• CrewAI — Python role-based agents vs. TypeScript DAG-based orchestration
• AutoGen — conversation-driven agents vs. task-graph coordination
• LangGraph — Python state-machine agents vs. TypeScript automatic task decomposition
• Mastra — full-stack AI framework vs. focused multi-agent orchestration library

FAQ

Q: Can I use different models for different agents? A: Yes. Each agent can be configured with a different LLM provider and model, allowing you to optimize cost and capability per task.

Q: How does it handle failures in subtasks? A: The coordinator can retry failed subtasks, reassign them to different agents, or adjust the plan based on partial results.

Q: Is it suitable for production use? A: The framework is designed for production use with proper error handling, retries, and observability hooks.

Q: Can I define my own task decomposition logic? A: Yes. You can override the default coordinator with custom planning logic while still using the DAG execution engine.

Sources

• https://github.com/open-multi-agent/open-multi-agent
• https://github.com/open-multi-agent/open-multi-agent/blob/main/README.md