★ OverviewIntermediate10 min

LangSmith: See Everything

Full observability for your agent: automatic tracing, run visualization, latency breakdowns, and debugging failed runs step by step.

Quick Reference

→Set LANGCHAIN_TRACING_V2=true and LANGCHAIN_API_KEY to enable automatic tracing — zero code changes required
→Every LLM call, tool invocation, and chain step is captured as a span in a hierarchical trace
→Use the LangSmith UI to inspect exact prompts sent to the LLM, token counts, and latency per step
→Tag runs with metadata (user_id, environment, version) for filtering and aggregation in dashboards
→LangSmith auto-captures input/output for every node in a LangGraph — including intermediate state

Zero-Config Tracing

Two env vars, full observability

LangSmith auto-traces every LLM call, tool invocation, and chain step with just two environment variables. No code changes, no SDK initialization, no decorators. Every LangChain and LangGraph execution is captured automatically.

Enable tracing — this is all you need

Every node execution becomes a span in a hierarchical trace. LangGraph nodes are auto-instrumented: you see the exact prompt sent to the LLM, the tool arguments, the state at each step, token counts, and latency per span — all without touching your application code.

Anatomy of a Trace

▸run_id — unique identifier for the entire agent invocation (the root span)
▸parent_run_id — links child spans to their parent, forming the trace tree
▸inputs/outputs — exact data flowing in and out of each step, including intermediate state
▸token counts — input_tokens, output_tokens, and total_tokens per LLM call
▸latency — wall-clock time for each span, broken down by LLM wait vs computation
▸error — full exception traceback if a step fails, pinpointing exactly where the agent broke

Metadata & Filtering

Tag runs with metadata for filtering and aggregation

Sign in to read this article

This is a premium article. Sign in with your Google account to continue.