TopolearnSeeing each step, state and model call is the fastest way to understand and fix agent behaviour.
LangSmith Studio is a visual IDE for agents. It draws your graph, lets you chat with or invoke it, shows the state after every node, and lets you rewind and fork (time travel) by clicking. LangSmith is LangChain's platform for tracing (recording every model call, tool call and step of a run), evaluation and deployment; Studio is part of it.
To use Studio locally:
1. Add a langgraph.json file to your project. It lists dependencies, a graphs map from a name to "./path/file.py:variable" (your compiled graph), and an optional env file. 2. Install the CLI: pip install -U "langgraph-cli[inmem]" (Python 3.11+). 3. Put a LangSmith API key in .env as LANGSMITH_API_KEY. 4. Run langgraph dev. It starts a local Agent Server (an API server that hosts your graphs, with in-memory persistence) on port 2024 and prints a Studio link.
The local server reloads on code changes, so you can edit a node and re-run from a past checkpoint.
Tracing without Studio: set LANGSMITH_TRACING=true and LANGSMITH_API_KEY in the environment, and every LangGraph and LangChain run is recorded in LangSmith, where you can inspect inputs, outputs, latency and token use per step. Tracing works for any Python script, not only langgraph dev.
Don't send sensitive production data to a tracing service without checking your organization's policy; LangSmith has options to hide inputs and outputs.
Example
{
"dependencies": ["."],
"graphs": {
"support_agent": "./src/support/graph.py:graph"
},
"env": ".env"
}Remember
- langgraph.json maps graph names to 'file.py:variable'
- langgraph dev runs a local Agent Server and links to Studio
- Studio shows the graph, per-step state, and time travel
- LANGSMITH_TRACING=true records runs in LangSmith from any script