Introduction

open-multi-agent is a TypeScript-native orchestration framework for Node.js backends. It lets specialist agents work as one team: for a non-trivial goal, runTeam() can turn the outcome into a task DAG, run independent work in parallel, and synthesize the results.

Start with the outcome, not a workflow diagram

Describe what the team should achieve. OMA can decide how to divide the work at runtime. If the task is simple or the graph is already known, you can choose a lighter or more explicit run mode instead.

Is OMA a fit?

Use OMA when

• A goal benefits from different specialist roles.
• Some work can run in parallel.
• The steps may change with the input.
• You want orchestration inside an existing Node.js backend.

Keep it simpler when

• One model call already solves the task.
• Every step and transition must stay fixed.
• You only need a provider SDK or chat UI.
• You do not want the cost of an extra planning call.

One run in four steps

1. Goal — your application supplies an outcome and a team of agents.
2. Plan — for a non-trivial runTeam() goal, the coordinator creates tasks, dependencies, and assignments. A simple goal may go directly to one agent.
3. Execute — the scheduler unblocks tasks as their dependencies finish and runs independent tasks in parallel.
4. Result — runTeam() synthesizes the task outputs. Individual agents can also return schema-validated structured data when you configure an outputSchema.

The architecture page shows the runtime components behind this flow.

Four core concepts

Agent

A role with a model, instructions, and an explicit set of tools.

Team

The agents available for one goal, with optional shared memory.

Task

A unit of work with an assignee and optional dependencies.

Orchestrator

The API that creates teams, schedules work, and returns run results.

What you can configure

Models & tools

Mix hosted or local providers, grant least-privilege tools, and connect external agents.

Orchestration

Choose automatic planning or an explicit graph, then add consensus, model routing, or plan replay when needed.

Reliable operation

Trace runs with observability, recover with checkpoints, control context, and measure quality with evaluation.

Choose a starting mode

• Use runAgent() for one agent and one focused task.
• Start with runTeam() when you know the goal but do not want to maintain the task graph.
• Use runTasks() when the graph and assignments must be explicit.

The Quick Start gives you a no-key first run in about five minutes. Then Choose a Run Mode explains the tradeoffs in detail.

Still comparing frameworks? See OMA alongside LangGraph, CrewAI, AutoGen, and the OpenAI Agents SDK on the comparison pages.