What does an unbounded retry loop cost?

Money and capacity, almost every time. Across 68 confirmed infinite-loop cases the same study found earlier, 95.6% drove what the authors call API cost exhaustion, repeated calls billed at full price with nothing to show for them, and the identical 95.6% also tied up model service capacity the same way a denial-of-service does, since every retry occupies a request slot another task could have used. Context window exhaustion followed in 27.9% of cases, where the loop’s own retries or accumulating messages filled the context until the agent could no longer function, and 7.4% burned through an external tool’s own rate limit on top of the model’s.

Those first two costs land together almost every time because they come from the same repeated call, not from two independent problems. A loop rarely picks one cost and stops there: it drains a budget and a shared resource in the same motion, the exact pattern Tessary’s cost_drift classifier is built to catch, since a loop with no bound shows up as spend on a call site climbing well past its own history before anyone notices the loop itself.