LangSmith is an observability and evaluation platform from the LangChain team that traces every step of an LLM or agent run, scores output quality against test datasets and LLM-as-judge evaluators, and tracks prompt versions and deployments, working with LangChain, LangGraph, or a custom stack built on any model provider.
Do I need to use LangChain to use LangSmith?
No. LangSmith is framework-agnostic by design: it traces runs built with LangChain or LangGraph, but also plugs directly into a custom stack calling a model provider’s SDK with no LangChain code in it at all.
What does LangSmith actually let a team do day to day?
Trace a run step by step to see exactly which prompt, tool call, or retrieval produced a given output, run AI agent evals against a labeled dataset before shipping a prompt change, and compare prompt versions against each other rather than guessing which one performs better. For anything running as an agentic workflow with several steps, tracing is usually the only practical way to find which one actually broke; without it, debugging a multi-step agentic AI system means reasoning backward from a bad final answer with no visibility into the steps that produced it.
Frequently asked questions
Do I need to use LangChain to use LangSmith?
No. LangSmith is framework-agnostic: it traces runs built with LangChain or LangGraph, but also plugs directly into a custom stack calling a model provider's SDK with no LangChain code involved at all.
What is the main thing LangSmith is used for day to day?
Tracing a run step by step to see exactly which prompt, tool call, or retrieval produced a given output, and running evals against a labeled dataset before shipping a change, rather than debugging a regression by guessing from the final answer alone.