← All field notes
CostApr 2025·6 min
Why per-step token caps beat per-call rate limits
Rate limits were designed for APIs, not autonomous agents. A looping agent can stay under every rate limit and still burn your budget by morning.
Rate limits don't stop loops
An agent stuck in a reasoning loop makes perfectly well-behaved requests — just thousands of them. Per-call limits never trip; the invoice does.
Budgets as a first-class primitive
Capsule tracks spend per agent, team, and org in real time. You set a hard ceiling, and when it's reached the run pauses before the next model call — not after.
Pausing safely
A paused run keeps its full state, so you can inspect, raise the cap, and resume — or let it stay parked. No silent overruns, no surprise bills.
Contain your agents with Capsule.Book a demo →