# Latency & performance

Where the time goes in a turn — response times, time-to-first-response, and turns per session.

The **Latency & performance** tab is about speed: how long your agent takes to respond, where that time goes within a turn, and how many turns a conversation runs. It's the view you open when the complaint is "the agent feels slow." Reach it at `/dashboard?tab=latency`.

## What it is

This tab breaks down your agent's timing. Where the [System performance](/docs/platform/performance.md) tab tells you *how much* traffic and *how many* errors, this one tells you *how fast* — the response-time picture across your traffic.

## The metrics

- **Time-to-first-response (TTFR)** — how long until the user sees the first token or response. This is the number users actually feel; a good TTFR keeps a conversation feeling responsive even when the full answer takes longer.
- **Total trace duration** — how long the whole turn takes end to end, across every LLM and tool call. The gap between TTFR and total duration is the work that happens *after* the user starts seeing output.
- **Turns per session** — how many back-and-forth exchanges a conversation runs. Read alongside the timing metrics, it tells you whether sessions are long because they're productive or because the agent isn't resolving things quickly.

## The timing diagram

A **timing diagram** visualizes where the time goes across a single turn — laying the phases of a turn out in sequence so you can see which part dominates. Use it to tell a slow *model* apart from a slow *tool call* apart from time spent between steps, without reading a raw span timeline.

## How to act on it

1. **Optimize TTFR first.** It's the latency users perceive. Getting the first token out sooner often beats making the whole turn faster.
2. **Decompose before you optimize.** Use the timing diagram to find the dominant phase — there's no point speeding up the model if a tool call owns most of the turn.
3. **Read turns-per-session as a quality signal.** Climbing turn counts can mean users are fighting to get an answer, not that they're more engaged.
4. **Drill into a slow session.** When a metric looks wrong, open the session and read its per-span timeline for the specific culprit — see [Sessions](/docs/platform/sessions.md).

## Availability

Any member of the tenant can view this tab. It's on by default — if you don't see it, your tenant has turned it off.

## See also

- [System performance](/docs/platform/performance.md) — volumes, error rates, and cost alongside these latency metrics.
- [Sessions](/docs/platform/sessions.md) — open a slow session and read its per-span timeline.