Agent operations: running the fleet.
Building an agent is a demo; running agents is a discipline. Once Researcher, Analyst, your Studio agents, and scheduled jobs are doing real work on real cadences, you've become an operations manager — this lesson is the ops manual.
01 The fleet mindset
One agent is a tool. Several — a Studio agent answering the team's policy questions, Researcher runs on a schedule, an intake flow triaging requests — are a fleet: they consume budget, produce output people rely on, and fail silently if unwatched. The shift that matters is treating them like staff you manage, not features you enabled. Staff get job descriptions, check-ins, and performance reviews. So do agents.
02 The agent register: one page, always current
For every running agent: Name · What it does (one line) · Trigger/schedule · Who consumes its output · What it can touch · Owner · Last reviewed. Ten minutes to create, and it answers the three questions that otherwise bite you at the worst time: what's running? who cares if it stops? and who do I call when it's weird?
03 Monitoring: agents fail quiet, not loud
04 The weekly fleet review (15 minutes)
Same three questions every week: What did each agent produce that someone actually used? What did any agent get wrong? What's one agent we should add, change, or kill? That last question keeps the fleet honest — agents accumulate like subscriptions, and an agent nobody would miss is pure risk surface plus cost. Retirement is a healthy outcome: log what it did, turn it off, keep the register accurate.
05 Cost discipline
Agent usage is metered in most setups — schedules multiply invocations, and enthusiastic triggers multiply them faster. The discipline: every agent's register line gets a rough monthly cost, reviewed against the 'someone actually used it' answer. The fleet earns its budget the way employees do — visibly, or not for long.
Build the register for everything currently running — yes, including the scheduled prompt you forgot about. Then hold your first 15-minute fleet review and kill or fix at least one thing.
Open Copilot →This week's challenge
Four weekly fleet reviews in a row — same three questions, register kept current, at least one add/change/kill decision per week. A month from now you'll have the thing almost no team has: an agent fleet that's actually managed.