7. See the analytics
One run tells you little; a hundred runs tell you where your delivery process actually leaks. Analytics in the sidebar aggregates them.
Across the top: a period, a project filter, and a person filter (Whole team by default). Below that, four views — Overview, Pipelines, Tasks, People.
Headline counters for the selected period: tasks completed, tasks in progress, runs finished, success rate.
The screen is careful about what it is counting — “1 run finished, 0 tasks marked done — these count different things”. Completed counts transitions into a done status inside the window, so a zero means no transitions landed there, not that nothing shipped.
Where the time goes
Section titled “Where the time goes”The breakdown that usually starts arguments, in a good way. Total run time split into:
- Agent — steps where an AI was working;
- Waiting for people — approvals and escalations sitting in someone’s queue;
- Checks — tests and automated reviews;
- Idle — everything else.
On a healthy small run it looks like Agent 28 s · 98%. On a real team, the
“waiting for people” slice is the one worth looking at: it is the cost of your
gates, measured instead of guessed. If it dwarfs the agent time, the problem is
not the model.
Returns
Section titled “Returns”How often work came back: runs with a return, and how many steps had to be redone. This is rework made visible — the pipeline transitions you drew for failure paths are what feed this number.
A rising return rate on one pipeline step usually means its acceptance criteria, or the prompt above it, is not good enough yet.
What the work cost in model calls — when those calls go through the platform. If your agent talks to its provider directly, this section honestly says so rather than showing a zero that looks like free work.
Where to go from here
Section titled “Where to go from here”You have run one task end to end: project, repository, pipeline, task, run, review, measurement. From here the usual next moves are to make the pipeline match how your team really works — more checks, an approval where it matters, a Set Status step so the board updates itself — and to bring a second person in so approvals have a real owner.