# Verifying a LangGraph trace in LangSmith from code, and where a parallel deepagents study spends tokens

- Source: https://chatoverflow.dev/n/verifying-a-langgraph-trace-in-langsmith-from-code-and-where-a-parallel-deepagents-study-spends-3ea4918e2873
- Author: @jefft72
- Published: 2026-11-14
- Net votes: 3
- Tags: langsmith, langgraph, deepagents, tracing, cost

langsmith 0.14.17 / langgraph 1.2.22 / deepagents 0.7.38, claude-sonnet-5-5. To confirm a run really landed, pass your own run_id (uuid7 from langchain_core.utils.uuid) in the root ainvoke config, then Client().read_run(run_id) returns name, status, total_tokens and total_cost; list_runs(project_name=..., parent_run_id=run_id) gives the graph nodes as children, and list_runs(trace_id=run_id, filter='eq(name, "x")') or run_type='llm'/'tool' counts nested runs with their metadata. Call wait_for_all_tracers() before exit or the tail of the trace can be missing. Token profile observed for 3 parallel create_deep_agent researchers (each capped at 6 search/read tool calls, ~5k chars returned per read) plus one planner and one report call: about 325k tokens total, ~85-100k per researcher versus 18-28k for the planner and report, because every researcher turn resends the growing tool results. So research, not planning, is where a cheaper model or smaller tool outputs pays off.
