Skip to main content

Compare session efficiency

Efficiency adds up the time your agents spent over a period and splits it into phases — exploring the code, writing the change, verifying it, waiting — so you can see where runs go slow or cost the most, and open the sessions behind the figures.

You need a role that can view sessions, and your organization's plan must include Session transcripts. Without it, Efficiency shows Not on your plan in place of its figures: org admins get a See plans link, everyone else reads Ask an admin to change the plan.

Not on your plan: your plan doesn't include session transcripts, with a See plans link

  1. Select Audit in the sidebar, then the Sessions tab, then Efficiency.
  2. Pick the period at the top right of Where the time goes: 7d, 14d or 30d. The dates under the heading are in UTC, followed by how many sessions they cover. The first scan of a period can take a while; the line at the bottom of the card says whether you are reading a Cached answer or a Fresh scan.
  3. Read the bar and its legend: each phase's share of the time and its total. Time spent awaiting a human is left out of the split — it is people's time, not the agent's — and named under the bar.
  4. In Per phase, compare the phases across their columns: Total, Share, Median / run, Sessions, Calls, Tokens in / out and Cost. A phase with a high Median / run is slow in a typical session, not just in one long one. Select Waiting to see what kind of waiting it was.
  5. Check where the effort was repeated: Gate churn shows how often the project's gate had to run more than once and how much time went on rework, and Duplicated work how often agents read the same thing twice.
  6. In the Sessions list, select a heading with an arrow, such as Total or Cost, to sort the sessions by it and find the worst ones.
  7. Select a session to open its task's sessions in the Sessions list, where you can see how each run ended.
  8. Optional: under Project memory, pick a Project (and a Persona if you like) to compare its sessions run with project memory on against those run with it off over the same period.
  9. Switch to another period with 7d, 14d or 30d and compare the same figures, such as Genuine exploration or a phase's Share, to see whether a change helped.