Dynamic and explicit orchestration
runTeam() builds a task DAG from a goal. runTasks() runs a graph you define, and runAgent() covers the single-agent case.
← All comparisons Enterprise support
// comparison
open-multi-agent vs Inngest AgentKit
Both build multi-agent systems in TypeScript. AgentKit routes a network of agents with deterministic, state-based logic on top of Inngest; open-multi-agent decomposes a goal into a task DAG at runtime.
Pick Inngest AgentKit if
You want deterministic, inspectable routing you control, and durable, replayable execution; and you’re happy to run on Inngest.
Pick open-multi-agent if
You want the plan built at runtime instead of hand-authored routing, no Inngest dependency, and a hard token budget.
01 at a glance
Side by side.
| Dimension | open-multi-agent | Inngest AgentKit |
|---|---|---|
| Language / runtime | TypeScript-native; embeds in any Node.js 18+ backend | TypeScript-native; pre-1.0 (0.13) |
| Orchestration model | Three modes: one agent, an explicit task DAG, or a goal decomposed by the coordinator at runtime | Multi-agent networks with deterministic, state-based routing; a router (code or model) picks the next agent |
| Runtime dependencies | 3 direct (Anthropic SDK, OpenAI SDK, Zod); extra providers and MCP are opt-in peers | 6 direct; runs on Inngest for durable, replayable execution |
| Mixed-model teams | Yes; each agent can use its own cloud or local model in one team | Yes; per-agent model via @inngest/ai adapters |
| Run-budget control | Token and estimated-USD ceilings through maxTokenBudget, or maxCostBudget with your estimateCost price table | No hard token cap; maxIter caps router iterations |
| Observability | TraceRecord v2 + TraceStore, an optional first-party OTel adapter, and an offline post-run Run Viewer | Run traces via the Inngest platform it runs on |
02 actual capabilities
What open-multi-agent includes.
OMA is more than goal decomposition and a small dependency count. These are current framework capabilities documented in the project README.
Deterministic control around agents
Inspect and approve plans, freeze and replay them as data, validate outputs with Zod, stream per agent, cancel runs, or add a proposer and judge consensus loop.
Dependency scheduling and recovery
The scheduler runs independent branches in parallel. Retries and checkpoints let an interrupted run resume without repeating completed tasks.
Production controls
Bound work with turn, token, estimated-cost, timeout, context, and loop controls. Tools are default-deny, and trace payloads redact secrets by default.
Your environment and your models
Run in your Node.js backend, locally, offline, or air-gapped. Mix cloud and local models, connect MCP tools, and bring external agents through ACP or process backends.
Inspect, trace, and evaluate
Stable run identity, TraceStore, and the offline DAG and Waterfall Viewer work without a hosted service. An optional OTel adapter and EvalSets connect runs to production telemetry and CI gates.
03 mechanism
How they differ.
AgentKit models work as a network of agents sharing state, with a router; code you write or a model you delegate to; deciding which agent runs next, capped by maxIter. It runs on Inngest, so execution is durable and replayable. open-multi-agent doesn’t ask you to author the routing: a coordinator decomposes the goal into a task DAG at runtime and parallelizes the independent nodes. AgentKit gives you explicit, deterministic control flow (and Inngest’s durability); OMA gives you a plan generated per goal and no orchestration service to run.
Where Inngest AgentKit fits
Choose AgentKit when you want deterministic, inspectable routing you author yourself and Inngest’s durable, replayable execution underneath; valuable when a run must survive restarts and every routing decision should be explicit and reproducible. It’s pre-1.0, so expect some churn, and it assumes Inngest in your stack.
Inngest AgentKit on GitHub↗Where open-multi-agent fits
open-multi-agent fits when you’d rather describe the goal than author the routing, and you want to stay dependency-light: the coordinator plans the task DAG at runtime, there’s no orchestration service to stand up, and maxTokenBudget gives a hard spend ceiling. Checkpoint/resume covers crash recovery at task granularity over any MemoryStore, without a separate durable-execution backend.
// Enterprise
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.
// Compare with another framework
open-multi-agent vs LangGraph → open-multi-agent vs CrewAI → open-multi-agent vs AutoGen → open-multi-agent vs OpenAI Agents SDK → open-multi-agent vs Mastra → open-multi-agent vs Vercel AI SDK → open-multi-agent vs VoltAgent → open-multi-agent vs LangChain → open-multi-agent vs LlamaIndex → open-multi-agent vs Pydantic AI → open-multi-agent vs Google ADK → open-multi-agent vs Semantic Kernel →