LangSmith Deployment vs Orchestrator Zero
oz0 records agent jobs running on your infrastructure. LangSmith combines agent deployment with tracing, evaluation and debugging. oz0 focuses on self-hosted operation across your machines, with outbound-only nodes, tenant rules and a model gateway.
What LangSmith Deployment is
LangSmith Deployment, formerly LangGraph Platform, is the hosting part of LangChain's LangSmith. Agents run on Agent Server, where queue workers pull runs from a durable task queue and write checkpoints to PostgreSQL. It runs on LangChain's cloud, as a hybrid with Agent Servers in your VPC, or self-hosted on Kubernetes with an Enterprise licence.
docs.langchain.com- Runs
- Self-hosted or managed
- Checked
- 27 October 2026
Side by side.
LangSmith Deployment's ratings come from its own pages. A column's name opens its definition and how Orchestrator Zero does it. How we rated
Where it runs
Any machine you own
Yes
DocsConditional: Enterprise plan
A standalone Agent Server runs in Docker on any machine, but running it yourself in production needs an Enterprise plan and its licence key.
Linux, macOS and Windows
Conditional: Docker on macOS and Windows
The production Agent Server is a Linux container image (Debian or Wolfi), so macOS and Windows run it only in Docker; langgraph dev runs natively anywhere but is for development and testing.
Self-hosted control plane
Yes
DocsConditional: Enterprise plan
The control plane can run in your own Kubernetes cluster alongside self-hosted LangSmith, but only on the Enterprise plan.
No inbound ports
Yes
DocsConditional: Server takes requests
The control plane never connects to the data plane, but the Agent Server is an API server that takes requests, and the default self-hosted layout runs work on that same host.
Fleet operations
Yes
DocsConditional: Deployments, no routing
The control plane manages deployments and shows their CPU and memory, but has no machine join, labels or routing of work to chosen hosts.
Durability
Durable runs
Yes: Checkpointed workflows
Agent Server persists checkpoints and can resume interrupted workflows. Checkpoint frequency and recovery depend on the graph's durability mode; interrupted work can run again.
Governance
Tool policy outside the model
Conditional: Middleware in agent code
The human-in-the-loop middleware checks each tool call against an allow, approve or reject policy, but the developer sets it in the agent's code, with no admin policy enforced by the deployment.
Approvals in chat
Conditional: Slack cards in beta only
Workflow interrupts persist while awaiting an API or Studio response. Managed Deep Agents adds Slack approval cards and is documented as a public beta on LangSmith Cloud.
Tool snapshots
Yes
DocsNot documented
Searched the LangSmith and LangChain docs for tool definition changes, tool poisoning and tool pinning; only server API version pinning is documented.
Tenants built in
Yes
DocsYes
Organisations hold workspaces, and each workspace is its own boundary for resources and access control.
Cost
Budgets that stop a run
Conditional: Caps in gateway beta
Spend policies in the LLM Gateway block any request that would push spend past a cap, per organisation, workspace, API key or user. The gateway is in beta; it runs on LangSmith Cloud, BYOC and self-hosted LangSmith.
Security
Provider keys off the workers
Yes
DocsConditional: Optional gateway, beta
The optional LLM Gateway, a beta, keeps provider keys in workspace secrets; by default the deployment's code holds the provider key in its environment.
Sandbox with an egress allowlist
Yes: LangSmith Sandboxes
LangSmith Sandboxes run agent code in isolated environments, and an allow_list on the egress proxy makes every destination not listed unreachable; it is generally available on LangSmith Cloud and BYOC and can be enabled self-hosted.
Audit and compliance
Tamper-evident audit trail
Conditional: Admin actions, Enterprise
Audit logs record administrative and configuration actions on the Enterprise plan and are called tamper-resistant, but no hash chain or verification is documented, and agent runs are kept as traces, which are not audit entries.
Erasure by key
Conditional: Delete traces by metadata
The API can delete traces selected by metadata, and custom encryption can use per-tenant keys. Built-in per-person keys covering histories and backups are not documented.
Retention with legal hold
Conditional: Retention, no legal hold
Admins set trace retention per workspace or project and TTLs for threads and checkpoints, but no legal hold is documented.
Openness
Model provider choice
Yes
DocsYes
Deployed agents use any LangChain chat model integration, which covers the major providers and local models through Ollama.
MCP and A2A
Yes
DocsYes
Agents call MCP tools through the MCP adapter, and every Agent Server exposes its agents over A2A and as MCP tools.
Where LangSmith Deployment is stronger.
Managed cloud or hybrid hosting
LangChain can host the control plane, Agent Servers and databases, or run only the control plane while your Agent Servers stay in your VPC.
docs.langchain.comSOC 2 Type II and HIPAA
LangChain keeps SOC 2 Type II, GDPR and HIPAA compliance with annual third-party audits and penetration tests.
docs.langchain.comSAML SSO and SCIM
Enterprise plans add SAML 2.0 single sign-on and SCIM user provisioning.
docs.langchain.comMany agent frameworks
Agent Server deploys LangGraph and LangChain agents and wraps Google ADK, Claude Agent SDK, Strands, CrewAI and AutoGen agents.
docs.langchain.com
Full in LangSmith Deployment, partial in oz0: Durable runs, Sandbox with an egress allowlist.
Differences to examine.
Compare the scope, conditions and editions in the source notes.
- Any machine you own
- Linux, macOS and Windows
- Self-hosted control plane
- No inbound ports
- Fleet operations
- Budgets that stop a run
- Provider keys off the workers
- Tamper-evident audit trail
- Erasure by key
Using both
A LangChain agent on LangSmith Deployment can call an Orchestrator Zero agent as an MCP tool. The two keep run state differently, so a run cannot move from one to the other.
Based on documentation; this pairing has not been tested.
Find your first task.
Explore an example or read the public Docs preview. Software access is coming soon.