What a fleet changes
The moment there are two devices, everything becomes a coordination problem: who runs what, who spends what, and who did what. Fleet management is just those three questions answered centrally — the mechanism is the same as multi-device claiming, applied with discipline.
Name everything
Every device gets a real name: office-mini, not MacBook-Pro-2023-2. Names appear in logs and audit entries, and “which machine did this?” should never take longer than one glance.
One policy, applied everywhere
- Write the fleet baseline once: allowed tools, paths, and secure mode defaults.
- Apply it to every agent profile — see central policy.
- Exceptions are per device and dated; review them monthly.
Budgets are fleet-wide
A fleet with ten small budgets still overspends ten times. Set one daily cap and per-task token budgets, then alert on the whole fleet. Per-device limits only make sense for a device with a known expensive job.
Read the audit as one stream
Every device writes to the same audit history. Skim it weekly for three things: denials that repeat, spend that drifts, and devices that have gone quiet — a quiet device is often a broken one, per the offline guide.