Trace the graph path
Record graph_id, run_id, thread_id, node_name and edge transitions so a failed answer can be replayed.
LangGraph Observability / Agent Trace
LangGraph Observability
LangGraph observability is the production layer that explains why an agent produced a result: which node ran, which tool was called, which checkpoint was restored, which retry happened, and which human review state blocked or approved the run.
Keep tool-call evidence
Capture safe tool input summaries, status, latency, retry_count and error_code without leaking secrets.
Connect checkpoints
Link checkpoint_id and review_status so human-in-the-loop recovery is auditable instead of invisible.
Route failures
Timeouts, parse errors and missing state should map to explicit troubleshooting pages and not disappear inside logs.
Minimal trace schema
{
"run_id": "run_2026_001",
"thread_id": "user_session_thread",
"node_name": "risk_review_worker",
"status": "retrying",
"latency_ms": 1840,
"retry_count": 1,
"checkpoint_id": "ckpt_42",
"review_status": "needs_human_review",
"safe_error_summary": "Tool timeout after policy-safe retry"
}