The $700-in-72-hours story has one root cause: the loop kept making API calls because nothing was counting. No budget control, no kill switch — just a loop and a credit card. The framework did not fail; the framework never had a meter in the first place. Three patterns actually stop the bleeding.
Ceiling per plan, not per account¶
Accounts aggregate; plans bound. Put the ceiling on the unit of work — the run, the campaign, the task — so a stuck loop exhausts its own plan, not your credit card. A per-account ceiling just means the loop has to work longer to do the same damage; a per-plan ceiling means one bad task can never take down the budget for everything else. This is the difference between "the account ran out of money" and "the plan ran out of budget": the first is a company-wide incident, the second is a contained event with a name, a run ID, and a receipt.
No budget controls? Proxy it.¶
If your framework has no meter, put one in front of it. Helicone's free tier meters 10k requests/mo; LangWatch's free tier tracks 200k events/mo. Cost of the meter: $0 — cheaper than one more incident. Route the framework's traffic through the proxy, set the ceiling there, and the framework's missing budget controls stop mattering. And the enforcement rule is non-negotiable: when the meter trips, the result must be blocked and recorded as blocked — not retried, not warned, not billed. A meter that warns is a dashboard with extra steps. A meter that blocks is a control.
Expiry is a spend limit¶
Every bounded plan expires. An agent workload with no end date is a subscription you forgot you bought — it keeps billing because nothing told it to stop. Give every plan a lifetime at creation, and let expiry be the backstop that catches the cases the ceiling misses: the ceiling someone set wrong, the meter someone forgot to wire up, the workload that was legitimate last month and abandoned this one.
The takeaway¶
Count the work, bound the work, and end the work. That is the whole spend-control posture: a meter that counts, a ceiling that blocks, an expiry that stops. We wrote up the full pattern — with the failure modes — in the free Governed Agent Mesh Playbook (https://ghostcorpnet.com/playbook.md), but the three patterns above are what keep the credit card safe. The $700 story never needed a better framework. It needed a meter.