Claude dynamic workflows, and open-multi-agent
In May 2026, Anthropic shipped dynamic workflows in Claude Code: the model plans and orchestrates the work at runtime. open-multi-agent makes the same bet — and because it runs inside your own backend, it also has to answer who approved a consequential action and what you can prove about the run afterwards.
Model-driven orchestration inside Claude Code. Claude writes its own orchestration scripts and fans out tens to hundreds of parallel subagents in a single session, checking its work before anything reaches you.
Read the announcement↗An MIT-licensed TypeScript library you install into your own backend. The model can plan the work at runtime here too — and consequential actions wait for a durable approval, while the run leaves a record you can verify offline.
Quick Start→The same bet: the model plans the work.
Both are model-driven. You don't wire a fixed graph up front — you hand over a goal and the model plans the work at runtime, decomposing it into steps that run in parallel and pulling the results back together. Claude's dynamic workflows do this inside Claude Code; open-multi-agent's coordinator does it in your backend. Same idea — so this page won't argue over which is more dynamic. The useful question is the one that appears when that orchestration becomes part of a product you ship: what do you have to answer for?
The difference is what you answer for.
They aren't the same kind of thing. Claude dynamic workflows are a capability inside Claude Code, orchestrating Claude subagents. open-multi-agent is a library you install into a TypeScript backend and point at any provider. Here is how the two line up.
| Dimension | Claude dynamic workflows | open-multi-agent |
|---|---|---|
| Where it runs | Inside Claude Code — CLI, desktop, and IDE | Your own Node.js backend — installed with npm, no hosted service to adopt |
| What it is | A capability of Claude Code | An open-source (MIT) library you embed |
| Models | Claude subagents | Any provider — OpenAI, Anthropic, Gemini, Bedrock, or any local / OpenAI-compatible model |
| The plan | Orchestration scripts Claude writes and runs in the session, checking its work before returning | A task DAG you can inspect and replay as data — planOnly, createPlanArtifact, runFromPlan |
| Before a consequential action | Claude checks its own work before anything reaches you | A gate can return { action: 'suspend' }: the request persists beside the checkpoint, bound to a SHA-256 hash of exactly what the reviewer saw, and the decision is atomic and first-wins |
| Afterwards | Orchestration runs inside a single Claude Code session | verifyRun() re-reads the journal cold and offline, checking each block still reproduces from its named source event — and the governance gate is judged on the execution receipt, never on agent answer text |
Composable, not just parallel.
These aren't mutually exclusive. open-multi-agent speaks the Agent Client Protocol (ACP), so an OMA team can drive external coding agents — including Claude Code itself — as one agent inside the team. The model-planned orchestration you get in Claude Code can become a single node in a larger, provider-neutral run that you own end to end. The permission boundary is worth reading before you rely on it: an ACP child runs outside OMA's per-call tool gate, filesystem sandbox, and egress policy.
Where open-multi-agent fits.
Reach for open-multi-agent when the orchestration has to live inside your own product and you have to answer for it. It is an MIT library you npm install into a Node.js backend, on any provider — cloud, local, or air-gapped. Declare governanceIntent: 'required' with requiredRoles and the run is judged on its execution receipt: which roles ran, in what order. That gate never reads agent answer text, so a model cannot satisfy it by writing that a reviewer approved. The framework makes no analytics, licence, update, or phone-home request of its own.
What the framework does not give you.
The suspend/decide/resume API and the durable record exist; the reviewer UI, the notification, and the escalation path do not — those are yours. There is no authentication, no RBAC, and no tenant model. The run journal is an audit trail, not tamper-evident storage: it has no hash chain, no signature, and no WORM backend, and verifyRun() checks the window it can still read rather than the whole run. Only in-memory and single-file stores ship; anything with cross-process guarantees you write yourself.
Taking this to production?
open-multi-agent is MIT-licensed and free to run yourself. When you need it delivered, integrated, or supported on a deadline, 元定义科技 (YuanASI) offers commercial delivery and support.
Looking for a framework alternative?
Weighing orchestration libraries against each other is a different question. See how open-multi-agent compares with LangGraph, Mastra, and the other frameworks.
All framework comparisons→