The slow agents are also the expensive ones.
For platform and SRE — latency and cost in the same view.
Average latency hides the tail. The agents that spike to ten seconds at p99, or cost the most per call, don't show up in a mean — and those are exactly the ones worth fixing.
Averages lie about the tail.
A healthy mean can sit on top of a long tail of slow, costly calls that your users feel and your budget absorbs.
The tail is invisible
p95 and p99 are where the pain lives, and a single average latency number erases both.
Erratic agents hide
An agent that's usually fast but occasionally spikes looks fine on paper and bad in production.
Speed is the wrong target
Optimizing for raw latency misses the real goal: the agents that are slow and expensive at the same time.
Tail latency and cost, ranked together.
p50, p95, p99 per agent
The full distribution, not a mean — so the tail that users actually feel is front and center.
Cost per call alongside
Latency and spend in one view, so the slow-and-expensive compounds rank to the top on their own.
Re-routing levers
Latency by model shows where a cheaper or faster model would do the job without losing the outcome.
The slowest individual calls
Drill into the specific outliers, so optimization starts from real evidence, not a hunch.
Optimize the agents that actually cost you.
Tail latency and cost, ranked together — not a flattering average.