LangSmith Tracing

Point LangSmith at your AIAgent with four environment variables, and every LLM call, tool call, and reasoning step traces automatically. No code changes.

At a glance

  • Environment variables only. No SDK code changes.

  • Traces every tool call, including collaboration tools, payment tools, and any custom_tools you added.

  • Complements, not replaces, console-based tracing. See Logging & Observability for that without a LangSmith account.

Setup

Add these to your .env file:

Variable

Value

LANGSMITH_TRACING

true

LANGSMITH_API_KEY

Your LangSmith API key. See the LangSmith docs for how to create one.

LANGSMITH_PROJECT

Any project name, created automatically on first trace.

LANGSMITH_ENDPOINT

https://api.smith.langchain.com

 
LANGSMITH_TRACING=true
LANGSMITH_API_KEY=your_langsmith_api_key
LANGSMITH_PROJECT=AgentConnect
LANGSMITH_ENDPOINT=https://api.smith.langchain.com

Set these before starting your process. Changing them after agents are already running has no effect.

What You’ll See

Every trace shows the complete path a request took, one agent run, expandable into every LLM and tool call inside it:

LangSmith project dashboard showing multiple AgentConnect traces

Project view: every agent run, listed as a trace.

Open a trace to see the step-by-step breakdown, LLM calls, tool executions, and the decisions in between:

Detailed view of a single trace showing the sequence of LLM calls and tool executions

Trace detail: the full sequence behind one response.

Every tool call is logged with its inputs and outputs, built-in collaboration tools (search_for_agents, send_collaboration_request), payment tools (native_transfer, erc20_transfer), and anything passed via custom_tools:

Detail of a tool call within a trace showing input arguments and the returned result

Tool call detail: exact inputs and outputs.

Failures are marked at the step where they happened, with the error message attached:

A LangSmith trace highlighting a failed step and the associated error message

A failed step, isolated in the trace.

Token counts and latency are tracked automatically on every LLM call and trace, useful for cost and performance tuning without adding any instrumentation yourself.

Troubleshooting

Nothing shows up in LangSmith

Confirm LANGSMITH_TRACING=true was set before the process started, and that LANGSMITH_API_KEY is valid for the project named in LANGSMITH_PROJECT.

Traces are incomplete or missing tool calls

Custom tools need clear names and docstrings to show up meaningfully in a trace; see AIAgent for tool-writing guidance. A trace reflects exactly what ran, if a tool was never called, it won’t appear.

Next Steps