Last year a dev watched an agent burn $700+ in 72 hours. No budget control, no kill switch — just a loop and a credit card.
Here is what most teams get wrong about agent governance: they buy observability and call it governance. Observability answers "what did it do." Governance answers "was it allowed to" — and stops it when the answer is no.
The cheapest control costs $0¶
Give every plan a spend ceiling and an expiry date, and fail closed when the ceiling hits. Blocked, recorded, never retried automatically. Then replay your worst incident in a sandbox and confirm the loop dies at its ceiling before it touches real money.
Warnings decay; blocks hold¶
Alerts at 50% and 80% are information — useful, but they are not the control. A warning is an invitation for the loop to finish. The block is the control. The run that gets blocked and recorded is a cheap lesson; the run that gets warned is a future incident.
From insight to enforcement¶
Observability is not worthless — it is the evidence layer that governance needs to make decisions. But without an enforcement layer on top (spend ceilings, identity-scoped grants, hard-gated approvals, hash-linked audit rows), the dashboard is a post-mortem tool, not a control. The Studio Edition playbook lays out the full six-policy, two-schema system with copy-paste templates. And the question to take back to your team: what would a runaway loop cost you before anything stopped it?