LangGraph is an open-source library for building agents as graphs of nodes with durable state. It is the control-flow layer, and a good one. Vocion is enterprise-grade AI you own: an open source platform with apps out of the box, running in your AWS account or ours.
The two are not strangers. Vocion's own agent loop is built on LangChain's
deepagents, which runs on LangGraph (@langchain/langgraph is a dependency
of @vocion/core; the choice is recorded in ADR 0001 in vocion-core). So
the real question is not "LangGraph or Vocion". It is whether your team wants
to build everything around the graph itself, or start from a platform that
already has.
Checked against vocion-core v5.2.0. Several rows go to
LangGraph, and they stay in.
When LangGraph alone is the right answer
When your engineers want custom control flow in Python or JavaScript, you already have a LangChain stack, you change the graph shape daily, or you want the larger ecosystem: more integrations, more answers online, more engineers who already know the API. LangGraph's community is far larger than Vocion's. If the job is one agent built and used by one or two engineers, a platform around it is likely more than you need.
What you build around the graph
LangGraph gives you the loop and its state. A team putting agents in front of the rest of the business usually ends up building the rest:
- Apps people use. A chat assistant, agents that do team jobs on a schedule, a place to see what they did. Vocion ships Workforce, Assistants, Software Factory and GTM.
- A deployment. Vocion runs in your own AWS account (provisioned with OpenTofu), on Vocion Cloud, or anywhere you run Node and Postgres 16.
- Connectors. Vocion registers 23 source connectors in core, including
HubSpot, Gmail, Google Drive, Slack, Jira, Notion, GitHub and S3
(
packages/core/src/libs/sources/registry.ts). - A record. Every
workspace:applywrites aworkspace_versionrow, and every tool call is stamped with theworkspace_shathat produced it, so an output traces back to the exact prompts behind it (README.md). - Cost and budgets. Per-agent budgets with soft and hard caps; a run over
a hard cap is refused before it starts (
services/BudgetService.ts).
None of this is impossible on LangGraph. It is work your team would own.
Side by side
| LangGraph | Vocion | |
|---|---|---|
| What it is | A library for agent control flow (docs) | A platform with apps, built on deepagents and LangGraph |
| Apps out of the box | None; you build them | Workforce, Assistants, Software Factory, GTM |
| Pausing for a person | interrupt(), resumed with a Command; needs "a durable checkpointer in production" (interrupts docs) | approve and ask workflow steps pause in a shared queue (docs/entities/workflow.md) |
| Surviving a restart | Your checkpointer | Workflow runs are durable on DBOS inside the app's own Postgres; a paused run waits through restarts and deploys (services/workflows/durableWorkflowRun.ts) |
| Where it runs | Your infrastructure, or LangSmith Deployment: Cloud, BYOC, Hybrid or Self-hosted (deployment options) | Your AWS account, Vocion Cloud, or self-hosted with Postgres 16 |
| Where agents are defined | Code in your project | YAML and Markdown in a git workspace, applied to Postgres (docs/workspace.md) |
| MCP | Supported via adapters (MCP docs) | MCP server over HTTP (multi-tenant) and local stdio (README.md) |
| Ecosystem | Large: the LangChain integration catalog and community | Smaller; 23 first-party source connectors |
| Control | Every node and edge is yours | Open source and configurable in YAML, but the loop is Vocion's |
| License | MIT (LICENSE) | MPL-2.0, with a commercial license available |
Keeping your own agent loop
Vocion can hand a run to a process it does not host. An agent with
harness.runsOn: external-worker queues a worker run; Vocion leases it,
records heartbeats and cost, and reaps it if the lease lapses, and whatever
the worker proposes lands in the same review queue as any other agent
(docs/entities/worker-run.md, behind VOCION_EXTERNAL_WORKERS=1).
Separately, packages/agent-runtime exposes one HTTP contract (POST /invocations streaming SSE, GET /ping) that can be hosted on AWS Bedrock
AgentCore. Wrapping an existing LangGraph graph in either is not a documented
integration today: treat it as wiring you would do yourself.
Which to pick
Pick LangGraph alone if you are a small engineering team that wants total control and the biggest ecosystem, and the people using the agents are the people who built them. Pick Vocion if you want apps your whole company can use this week, running in an account you own, with the agents and their history kept in your git and your Postgres.
Related
- Compare approaches: closed SaaS AI, building on a framework, or Vocion.
- Workflows: the
approveandasksteps. - Bring your own agent runtime: the HTTP contract.
- Enterprise: what ships today and what is on the roadmap.