Routing, latency, spend and quality in one view.
Where the traffic went, what it cost, how fast it came back, and whether the router’s choices are holding up.
- Model mix over time — watch adaptive routing shift as your prompt distribution changes.
- Spend by key, workspace, agent and model, so a surprise line item always has an owner.
- Agent tool-call rate spikes flagged against learned hour-of-week baselines rather than a fixed threshold.
Spikes measured against your own normal
Tool-call rate anomalies are scored against a learned hour-of-week baseline, so Monday 9am traffic is never compared to Sunday 3am. Baselines are per workspace, which means there is no global threshold to tune, and detections surface alongside the firewall activity they came from.
Watch the mix move.
Your prompt distribution changes, and adaptive routing changes with it. Insights shows that happening — which models are taking the work this week, what they cost, how fast they came back, and whether quality held while the mix shifted.
Surprises arrive early.
Tool-call rate spikes are scored against a baseline learned per workspace, per hour of the week — so Monday at 9am is compared to other Mondays at 9am, not to Sunday at 3am. You hear about it as a signal, while it is still a signal.