# Decide whether a node earns its keep

Read one node's page in the Profiler, from the quality it contributes against traces that skipped it to its latency percentiles, its worst answers and where its tokens go.

A node's page answers one question: is this node pulling its weight? It shows what the node costs in time and money, and what the answers look like when it is in the loop against when it is not.

**Who can use it:** the **Builder** role or higher, in a tenant licensed for agents.

## Opening one

A node page opens from a row in the Overview's per-node breakdown, or from **Investigate node** on an SLO group in [Issues](/agents/profiler-issues). The **Per-node** tab in the tab strip is never clickable, and the note under the strip says where to open one instead.

## Quality contribution

![A Profiler node page for the summarizer node: a quality contribution card comparing the judge score with and without the node, and five figures for p50, p95 and p99 latency, calls with average output tokens, and spend.](https://docs.katonic.ai/img/agents/profiler-node/profiler-node.png)
*A node's page asks whether the node earns its keep: what answers look like with it in the loop against without it, and what it costs in latency and spend.*

The lead card is an A and B comparison. The platform finds traces that ran through this node and traces that skipped it, scores both with the judge, and reports the mean of each with the difference and a confidence figure.

The verdict is one of two sentences. Either the node measurably improves answers, by the stated amount, and the latency and cost it adds are the price of that uplift; or answers are as good or better without it, and it should be considered for removal.

The comparison needs a cohort that skipped the node, which exists only where something routes around it. Without one, the card says per-node quality contribution needs per-node judge scores and tells you to enable auto-eval and let evaluated traces accumulate, rather than showing a comparison it cannot make.

## What the node costs

Five figures follow: **p50**, **p95** and **p99** latency, the last marked as the tail and counting the traces above five seconds; the call count with the average output tokens per call, which reads as a dash rather than zero when nothing was measured; and the spend in the window with its share of the agent's spend.

**Latency over time** plots the three percentiles across the window, marks the peak with its time, gives each series' mean in the legend and names the peak window. A window with no samples says so inside the chart.

## The worst answers

![A table headed Worst answers listing traces with their questions and answers, judge scores and the judge's notes, beside a latency distribution histogram for the node.](https://docs.katonic.ai/img/agents/profiler-node/profiler-node-worst.png)
*Worst answers lists the traces through this node with the lowest judge scores, with the judge's notes; a row opens the trace.*

**Worst answers** lists the six traces through this node with the lowest judge scores: the trace, the question and answer, the score and the judge's notes. A row opens [that trace](/agents/profiler-trace-detail).

**Latency distribution** buckets the node's calls across the window.

## Actions

**Tune this node** and **Ablation study** both open a new Optimizer run carrying the agent and this node, and the ablation carries its mode as well, so you do not pick them again.

## Troubleshooting

- **No data for node yet.** The node exists but has too little traffic. The page says the numbers start to fill in after about fifty completed traces.
- **Couldn't load the node.** The profiler service is unavailable, or the window is empty; reload or switch the range.
- The figures on this page carry no change arrows against the previous window, and there is no breakdown of where a call's input tokens go. Neither is in what the platform returns for a node.

## Related

- [Read the Overview](/agents/profiler-overview)
- [Read one trace in full](/agents/profiler-trace-detail)
- [Optimizer](/agents/optimizer)

Source: https://docs.katonic.ai/agents/profiler-node
