LangSmith.
Tracing, evals, and observability — the default observability tool for agent stacks.
What it is.
LangChain's observability product. Traces every agent invocation as a structured DAG, supports both LLM-judge and human-labelled evals, and exposes cost-and-latency dashboards. Used by most production LangChain and LangGraph deployments and increasingly by non-LangChain stacks.
Where it fits.
Once an agent stack crosses roughly 10K daily invocations, the team needs structured observability. LangSmith is the default — even teams that do not run LangChain in production often run LangSmith for the eval and trace surface.
- Best-in-class agent trace visualisation
- Evals framework with LLM-judge and human-label workflows
- Self-hosted option for regulated buyers
- Pricing climbs quickly with trace volume
- Native LangChain bias — non-LangChain integration requires more work
Frequently asked.
Do I need to use LangChain to use LangSmith?
No. LangSmith works with any framework via OpenTelemetry-style instrumentation. Native LangChain integration is the fastest path to first trace.
What is the difference between LangSmith and Weights & Biases?
LangSmith is agent-trace-first; W&B is experiment-and-training-first. Many teams use both, scoped to different parts of the stack.
Is LangSmith self-hostable?
Yes via the Enterprise tier. Standard deployment in regulated industries that cannot send traces to a third-party SaaS.
