Paying for Agents: Cost Control for Agent Fleets
AI
Agents can spend at programmatic speed. The controls that keep them affordable: route by task, reuse work, cap with budgets, and watch headroom as a metric.
The first time a system I built ran a fleet of agents on its own, the thing that surprised me was not what they did. It was what they cost. Each agent, left to its own loop, will call a model as many times as it decides it needs to, and a hundred agents doing that in parallel is a spending machine with no brakes unless you built the brakes in. Nobody is standing there watching the meter. The meter is running at the speed of software, and the only thing between you and a shocking invoice is the design.
TL;DR: Agent fleets spend money at programmatic speed, so cost has to be a design property, not an afterthought. Four habits carry most of it: route each task to the cheapest model that can actually do it, reuse work you already paid for instead of redoing it, put hard ceilings in the path so a runaway stops itself, and watch the headroom between what you are spending and what you have decided to allow. Do that and cost stops being a surprise and becomes a dial.
Agents spend at programmatic speed
A person using a model spends slowly, one thoughtful request at a time. An agent does not. It works in a loop, and a fleet of them works in many loops at once, and none of them pause to consider the bill. This is not a reason to fear agents; it is a reason to respect that their spending has a different shape than human usage. The cost does not creep up gently. It can spike the moment something loops more than you expected, which means the controls cannot be something you check weekly. They have to be in the path of every call, working at the same speed the spending does.
Route each task to the cheapest capable model
The single biggest lever is to stop sending everything to the strongest model. Most of the work a real system does is easy, and a smaller, cheaper, faster model handles it perfectly well. A minority is genuinely hard and deserves the expensive one. Routing by the task, not by habit, means the easy majority pays the cheap price and only the hard few pay the premium. The trick is honesty about which is which, and a willingness to send the easy work to the humble model even though the strong one would also, of course, succeed. Paying premium prices for easy answers is the most common way these systems waste money.

Reuse the work you already paid for
The second lever is to stop paying twice. Agents ask overlapping questions, retrieve the same documents, and re-derive the same intermediate results constantly. Much of that can be remembered instead of recomputed. When the same request comes around again, or a near-identical one, the cheapest possible answer is the one you already have. Caching is unglamorous and it quietly removes a large slice of spend, because in any real workload a surprising fraction of the work is not new. The discipline is to notice how much of what your agents do is repetition wearing a slightly different hat, and to pay for it once.
Budgets and ceilings, not surprises
Levers reduce cost; ceilings bound it. A budget that only exists as a number on a dashboard is not a control, it is a description of what already happened. A real ceiling lives in the path: when an agent, or a team, or the whole fleet crosses the line you set, the system slows or stops on its own, before the overspend, not after. This feels aggressive until the first time it saves you from a loop that would have run all night. The point of a ceiling is not to be hit; it is to make the worst case bounded instead of open-ended, so a bug becomes an annoyance rather than an event.

Headroom is a first-class metric
The number I watch is not just what we spent; it is the gap between what we are spending and what we have decided to allow. That gap is headroom, and it is the early warning the raw spend number never gives you. Plenty of headroom means the fleet can grow and absorb a bad day. Headroom quietly shrinking means trouble is coming before any ceiling is hit. Treating headroom as a first-class thing to watch turns cost from a rear-view mirror into a windshield. You stop asking what did we spend and start asking how much room do we have left, which is the question that actually lets you steer.
The takeaway
An agent fleet is a fleet of independent spenders, so affordability is architecture, not luck. Route each task to the cheapest model that can do it. Reuse the work you already paid for. Put ceilings in the path so a runaway stops itself. And watch headroom, the room between your spend and your limit, as a real metric rather than reading the bill after the fact. None of this makes the agents less capable. It makes them something you can afford to run at scale, which is the only kind of system that gets to keep running.