Choosing between control-plane-managed and standalone Agent Server deployments
Last updated: November 3, 2026
Overview
Self-hosted LangSmith supports two architectures for running agents: a control-plane-managed LangSmith Deployment, and a standalone Agent Server that you deploy and manage yourself. If you're setting up agents across multiple clusters or multiple namespaces, the deprecation warning on the "Configure additional data planes" docs section can make it unclear which path to standardize on. This article covers the scope of that warning and the tradeoffs of each architecture.
What the deprecation warning covers
The warning applies only to the "additional data planes" mechanism: configuring extra control-plane-managed data planes in other Kubernetes clusters, or in a different namespace within the same cluster, wired back to a shared control plane through the hybrid/listener setup. That specific path is not recommended going forward and is planned for deprecation.
It does not apply to the primary, single in-cluster control-plane-managed LangSmith Deployment architecture. That setup is unaffected and remains fully supported, with no announced timeline for removal.
Recommended architecture for multi-cluster/multi-namespace
For multi-cluster or multi-namespace deployments, use standalone Agent Servers instead of additional control-plane-managed data planes. A standalone Agent Server is a production-ready Agent Server you deploy and run independently: you own its lifecycle (deployment, scaling, CI/CD), and it optionally reports traces and evaluations to a LangSmith instance (self-hosted or SaaS) via the LANGSMITH_ENDPOINT environment variable and an API key.
Because the integration with LangSmith is limited to tracing, a standalone Agent Server is not coupled to the control plane's version. There's no minimum LangSmith platform version to keep in sync between a standalone server and your LangSmith instance.
For production standalone deployments, use Kubernetes with the maintained LangSmith Helm chart. Docker and Docker Compose are intended for local development and testing, not production.
Context Hub access from standalone Agent Servers
Context Hub works the same for agents on standalone Agent Servers as it does for control-plane-managed agents. Context Hub is a feature of the LangSmith instance itself, and agents reach it through the LangSmith SDK using the same endpoint and API key used for tracing. Whether an agent runs under control-plane management has no effect on Context Hub access.
Deployment discoverability and governance tradeoff
LangSmith does not detect or list agents running on standalone Agent Servers in the Deployments list; only control-plane-managed deployments in the main cluster show up there. For standalone Agent Servers, only the traces they send to their configured tracing project are visible in LangSmith.
If you need fleet-wide deployment discoverability or governance from the LangSmith UI, factor this in before choosing standalone Agent Servers for all of your multi-cluster or multi-namespace agents.