Why agents need a budget, not a dashboard

A dashboard tells you what your agents spent. A budget decides what they are allowed to spend. When the thing spending your money is an autonomous loop that runs at 3am, only the second one helps.

Why is a dashboard not enough?

Dashboards are built for a person at a console reviewing a bill after the fact. That model works when the caller is a human who makes a few decisions a day. It breaks when the caller is an agent: fast, literal, running in parallel and sometimes wrong.

An agent that gets stuck in a retry loop does not get bored. It does not notice that the last forty attempts failed the same way. By the time a chart turns red and someone looks at it, the money is gone. The dashboard was accurate. It just was not in the loop.

Most agents today run on a VPS with a cron job and a .env file holding a model key with no limit on it. Nothing between the agent and the provider says stop.

What does a real budget look like?

A useful budget has three properties:

1. It is in dollars. People reason about money, not vCPU-hours or tokens. "This agent may spend $30 a month and $0.50 per run" is a sentence anyone can approve.
2. It is enforced on the call path. Every model call and tool call is checked before it happens, not reconciled afterwards.
3. It fails safe. At the cap, work pauses with a clear reason, and nothing is lost.

In Heighliner, every fleet has a monthly cap and a per-run cap in its manifest. Model calls go through a gateway that attaches your own key (from any provider or OpenRouter), prices the call from a per-model table and spends from the fleet's budget. Tokens are tracked so the budget works, and never marked up.

How do leases keep the budget fast?

Checking a central ledger on every call would make every call slow. So the gateway does not ask the ledger each time. It takes a small budget lease, $1 by default, and spends from it locally. When the lease runs low, it asks for another one. Unspent money goes back when a lease closes or expires.

That design gives a precise guarantee: the hot path never waits on the database, and the worst overrun is one lease. Warnings go out at 50, 75 and 90 percent. At 100 percent no lease is granted, the call returns a budget_exhausted error that says how to fix it, and the agents pause until you raise the cap.

Does a budget make agents less useful?

The opposite. The reason people do not leave agents running overnight is that nothing stops a bad night from becoming an expensive one. A hard cap is what makes it reasonable to say yes to "run this every hour", to give an agent a schedule, or to let a fleet scale out on a research task.

The budget also becomes the one setting that matters. Instead of tuning instance sizes, you say what the work is worth and let the platform keep it inside that.

Where does the dashboard fit, then?

You still want to see spend. Heighliner shows it per fleet and per run, and every run ends with a receipt. But the dashboard is for understanding, not for safety. Safety comes from the cap, the approval gate for actions that cannot be undone, and the sandbox each agent runs in.

If you run agents on a cron job today, see how Heighliner compares with a VPS and cron.