What is Cost Creep?
Operations (LLMOps & AgentOps)The drift type where call volume, context size, and retry counts grow gradually until the invoice discovers what the dashboards did not. A pattern that cost cents in testing costs dollars at scale.
Why It Matters
Cost creep is the drift type that arrives on an invoice. A conversation pattern that cost cents in testing costs dollars at scale, and because each individual call looks normal, nothing alerts until the billing cycle closes. The DN42 incident is the infrastructure version: an agent asked to index a hobbyist network provisioned five 48-vCPU instances at roughly $270 a day, and discovery came from credit card charges. The $380 incident is the token version: a support agent hit a rate-limited tool and re-planned for forty minutes after the user walked away. Both systems had dashboards. Nobody was watching them. Dashboards are witnesses; budgets are brakes.
The Three Drivers
Longer reasoning paths. As the agent meets more diverse inputs, it finds longer chains, and each extra step is extra tokens.
Retry compounding. Tool failures trigger retries, and retries multiply: the failed call, the re-plan, and the repeated attempt all spend.
Context growth. Every turn sends the accumulated context back to the model, so turn fifty costs more than turn five even when the work is identical.
Where It Breaks
Cost creep survives because each cause is a feature: longer reasoning looks like thoroughness, retries look like resilience, and context growth looks like continuity. The break is measuring spend only at the aggregate. Without per-step cost attribution, the trend has no signature to alert on, and without cost bounds there is nothing between a creeping pattern and an unbounded bill.
How Flytebit Handles It
Cost creep maps to the monthly cost review in our maintenance cadence, with attribution per run and bounds enforced before calls execute. The DN42 and $380 post-mortems are dissected in Operating Agentic AI Systems, and the cost layer is part of our LLMOps work.