Insights

How to read the Insights page in paper console — spend, usage, and session health across your team, and how to turn the aggregates into sessions worth opening.

The Insights page turns your team’s captured session history into a read-out of spend and run health. Where the Sessions list answers “what happened in this run,” Insights answers the questions that come after a few weeks of real usage: what is all this agent work costing, where is the time and money actually going, and is it producing finished work?

Sample Insights dashboard: a stat strip with total spend, tokens used, agent time, tool calls, and success rate; a spend-over-time chart with one day pinned to show its top sessions; and a suggestions column flagging a cost outlier and spend on unfinished runs

What’s in Insights

Everything on the page is scoped to a time window you pick — the last 7 or 30 days — and every aggregate links back to the real sessions behind it.

The stat strip

Five topline numbers for the window: total spend, tokens used, agent time, tool calls, and success rate. Each comes with a per-session average and a delta against the previous period of the same length.

The absolute numbers tell you scale; the deltas tell you direction. Spend up 200% is a very different story depending on whether success rate held — one is a team adopting agents for more work, the other is money going into runs that don’t finish.

Spend over time

A daily cost chart for the window. The shape is the point: steady daily spend reads differently from one enormous spike. Click a day to pin its most expensive sessions and jump straight into any of them — the fastest route from “what happened Monday?” to the actual run.

Spend by model

Total cost per model, with session counts. This is where you sanity-check your model mix: if the most expensive model is absorbing routine work that a cheaper one handles fine, it shows up here first.

Longest and most expensive sessions

A ranked list you can flip between longest (by turn count) and most expensive. Long isn’t automatically bad — some work is genuinely deep — but a high turn count is often the first symptom of an agent looping, retrying, or grinding against missing context.

Suggestions

Prioritized observations the console pulls out for you: a session that cost several times your median, spend that went to runs that failed or were abandoned, and similar anomalies. Each suggestion links to the session it’s about, so the next step is always “open it and read the chain.” When nothing crosses a threshold, you get a calm all-clear.

Sample Insights breakdowns: spend by model shown as horizontal bars with per-model totals and session counts, next to a list of the longest sessions with turn counts, authors, and costs

Understanding the numbers

A few things to keep in mind when reading the page:

  • Insights only sees what you capture. Every number comes from sessions run through the paper CLI. If half the team runs agents outside of paper start, the dashboard undercounts by exactly that much — the fix is capture coverage, not the chart.
  • Deltas compare like with like. Each stat is measured against the previous period of the same length, so “up 24%” on the 7-day view means versus the 7 days before it.
  • Averages are per-session. The small figures under each stat (average spend, average duration, average tool calls) describe a typical run, which is often more actionable than the total.
  • Aggregates are entry points, not conclusions. The chart tells you Monday was expensive; only the session detail page tells you why. When a number surprises you, click through before drawing a conclusion.

Tips for analyzing

Start with the deltasTrends matter more than totals. Rising spend with steady success rate is adoption; rising spend with falling success rate is waste worth investigating.
Chase outliers firstOne session at many times your median usually explains more of the bill than everything else combined. Suggestions and chart peaks take you straight to it.
Read long sessions for loopsOpen the highest turn-count sessions and skim the chain. Repeated tool calls and retries point at missing context or a prompt worth fixing once, for everyone.
Match models to tasksUse spend by model to spot routine work running on your most expensive model — often the cheapest optimization available.
Treat failed spend as a signalMoney on runs that didn’t finish is a workflow smell, not just a cost. The failed sessions themselves show whether it’s flaky tooling, bad prompts, or abandoned experiments.
Zoom out before reactingFlip between 7-day and 30-day views to separate a one-off spike from a real trend before changing how the team works.

Frequently asked questions

Why is my Insights page empty?+
Insights is built entirely from captured sessions. If you haven't run an agent through the paper CLI yet — or no sessions fall inside the selected time window — there is nothing to chart. Run an agent with paper start and the page fills in as sessions are captured. See the Quickstart.
What does the Insights page show?+
Aggregate numbers for a chosen window — total spend, tokens used, agent time, tool calls, and success rate — plus a daily spend chart, spend broken down by model, your longest and most expensive sessions, and suggestions that flag cost outliers and failed runs.
How are the percentage changes calculated?+
Each stat is compared against the previous period of the same length. The 7-day view compares against the 7 days before it; the 30-day view compares against the previous 30.
Copied to clipboard