Agent team capacity and measured concurrency
New agent sandboxes size their CPU and memory for the configured worker ceiling, up to 4 vCPU and 4 GiB for 32 workers. Auto now receives that capacity through the server adapter.
The Team view and CLI show peak overlapping model calls measured from completed invocations. Queued tasks and planned team size are not counted as proof of concurrent execution. A persistent project budget can cap compute and managed inference together across rehearsals.
Agents can pass environment settings, project Secret references and an explicit outbound network policy to their tool sandbox, so teams can use scoped GitHub and other service access.