See how an agent decided
For a beta user: read the turn and tool-call counts on a finished run's page where the runtime serves them to you, and know which reads of how an agent decided this page does not cover.
For a beta user who wants to know how an agent reached its answer. After this page you can find the counts line on a finished run's page where the runtime serves it to you, and know which reads of how an agent decided this page does not cover. Not here: judging a run that went wrong (Judge a run you did not like).
Agentic introspection is family 22, which reads four bars evidenced in the census at revision 3ec0e4f4, 2026-10-20. The cloud manifest claims all four, and the census finds them evidenced. The reachable bar reads under the older census rule, not the newer one; the measured and natural-language bars read under the newer one. The census console bar has no reading key.
As of 2026-10-21, the cloud SDK and command line name one method for this family: runs.turns ("the model and tool turns a run took"), with the command-line verb runs turns. They name none for a run's durable facts. The cloud SDK is also built from source until its first publish. This page does not show that call, because the runtime serves that read only to the engineering audience (below).
Find the counts
Open Workflows in the left rail, open its All runs group, and open a finished run that the runtime still holds. Near the run's status is one line of counts, such as "N of M turns · N of M tool calls". When the run had any refusals, the line ends with "N refused and re-prompted".
The line shows counts: no grade, no money and no duration are drawn there. It is the run's score read, not one of this family's own reads.
The runtime serves that read to you in your own personal workspace and, in an org, to an owner or an admin where your org has them. For a person it is not served to, the line is absent, with no message. It is also absent when the runtime holds no score for the run, and on a run shown from its kept record after the runtime no longer holds it.
The turns
On a run the runtime still holds, the page draws a finished run's turns when the runtime offers that read to you. The runtime serves them only to a person on the platform's own engineer list, and no plan or organization role reaches that list. For the owner of a run who is not on that list, the page does not ask for the turns and draws nothing there. If it does ask and the runtime refuses the read as withheld, the page also draws nothing there and shows no error. A run shown from its kept record has no turns.
What this page does not cover
Dated 2026-10-21. None is a step on this page.
- The durable run facts. The runtime serves a run's durable fact on its own read, to the person whose run it is: the run id, when it settled, which clock that is on, and the build revision. It serves other fields to wider audiences. The run page does not draw it.
- Replan rounds, rerank turns and lineage. The contract capture this site pins marks these three as not served, and the runtime's own notes say they are not coming: a replan round carries no run key and none is written, a rerank read would return every tenant's reranks, and a lineage walk runs over every tenant's records.
- Failure, load and scenario checks. The census finds no instrument or receipt for this family for any of the three.
- Driven live. Everything on this page is read from the console source and the cloud code at the cited revisions. None of it was driven against a deployed console for this page.
What it refuses
The turns read and the facts read are each scoped to the signed-in person in the current workspace. A run that is not yours answers the same not-found as a run that does not exist.
The census tenancy check for this family reads PARTIAL: a static scope audit, and no cross-tenant probe receipt.