Token spend on agentic work is bursty and hard to predict: an agent that reads the wrong subtree can burn an hour's budget before producing anything. A cap converts an unbounded risk into a known one and forces the interesting question, which is what the task is worth.

Budgets also shape behaviour usefully. An agent told it has a limited budget prioritizes differently — it searches before it reads, and summarizes instead of quoting — which is often the behaviour you wanted regardless of cost.