Engineering · Performance

See how fast each agent responds and what each call costs — and where slow and expensive collide.

The problem

Speed and spend live in different tools, so the one combination that actually hurts — an agent that is both slow and costly — never shows up. Tail-latency outliers users feel go unnoticed until someone complains.

The value

One view ranks every agent by the latency its users actually feel (p95, not averages) against its $/call. The slow-and-expensive outliers surface first, so you fix the responses that hurt most and trim the spend that buys the least.

6.8s

worst p95 (Codex CLI)

1.1s

best p95 (data-pipeline-agent)

agents profiled