LangSmith · Traces view
LangSmith solved the hardest problems in LLM observability. This rebuild keeps the exact columns and layout — and elevates the surface to match the depth of the tooling.
| Run | Input | Start | Latency | Tokens | Cost | Status |
|---|---|---|---|---|---|---|
agent_executor 1f0a-8c21 | Find Q1 churn drivers across enterprise accounts | 09:41:22 | 8.41s | 14,382 | $0.0861 | Success |
retrieval_qa 1f0a-8b9e | What does clause 7.2 say about SLA credits? | 09:40:51 | 1.24s | 3,210 | $0.0192 | Success |
extract_entities 1f0a-8b41 | Parse vendor invoices from the shared inbox | 09:39:37 | 620ms | 1,847 | $0.0110 | Error |
ChatAnthropic 1f0a-8ad2 | Summarize the incident postmortem for exec readout | 09:38:04 | 2.13s | 5,904 | $0.0354 | Success |
rerank_documents 1f0a-8a67 | Top 20 support tickets mentioning latency | 09:37:48 | — | — | — | Pending |
agent_executor 1f0a-89f3 | Draft renewal email for at-risk accounts | 09:35:12 | 6.87s | 11,205 | $0.0673 | Success |
sql_generator 1f0a-8988 | Weekly active teams by plan tier, last 90 days | 09:33:40 | 940ms | 2,418 | $0.0145 | Success |
retrieval_qa 1f0a-8901 | Which regions are covered by the EU data addendum? | 09:31:09 | 1.58s | 3,877 | $0.0233 | Error |
classify_intent 1f0a-88b7 | Route inbound: 'need SSO enabled before Friday' | 09:29:55 | 310ms | 692 | $0.0041 | Success |
Rebuilt by Kyre — same layout, same data, same information architecture. Only the execution changed.
What changed — and why it matters
Two runs out of 1,204 failed, and they were buried in uniform gray. I made them the first thing you see — because during an incident, finding the failure is the job. That's a decision about what matters, not a color choice.
Latency and tokens compete for attention in every row. I decided latency wins — it's what predicts a bad user experience. Showing both equally is data; knowing which one you act on is judgment. The inline bar makes the slow run visible from across the room.
Loading, empty, error — most dashboards treat these as edge cases. They're a third of what users actually see. Each state now says what happened and what to do next, so a blank filter result never reads as a broken page.