Agent Budget Ceiling

Budgets designed for human users assume human pacing. An agent issues requests at machine rate and can exhaust a monthly allocation in an afternoon, which makes per-user ceilings an inadequate control once automation is involved. An agent ceiling is scoped to the process and the task, with defined behaviour at the limit — halt, degrade to cached responses, or escalate for authorization. Setting it requires knowing what a normal run costs, which requires metering before automation rather than after.

Because consumption is metered per execution and per skill in Centralpoint, a ceiling can be set against measured behaviour rather than estimated. Governed cached answers provide the degradation path: an agent approaching its limit can be served from the local index rather than halted outright, which keeps a long-running task alive without further model spend.