Run & approval
What a single agent execution records, and where a human stands in the way.
A run is one agent execution, scoped to a workspace, application, environment, tenant and end user. It records its model steps, the events it emitted, and the tokens it consumed against the ceiling the environment set.
Inspecting a run
Run activity lists executions for a bound tenant. Opening one shows its status, token usage, the model steps behind it with their individual input and output counts, and the ordered event timeline. A failed run carries a failure code.
Approvals
Read-only context can flow across systems. Anything consequential stops at a checkpoint, and the approval is bound to that exact proposal — approving does not grant standing permission for a similar action later.
Audit
Every recorded action in a workspace lands in the audit trail with its outcome, retained for the period the environment specifies. It is an append-only record, not a view you can edit.