Split an agent into a planner, a tool-caller and a policy checker and you are supposed to get a more careful agent. Charged to the same per-conversation budget of generated tokens as the single agent it replaced, over 7,950 policy-bound customer conversations, the split scored lower in 12 of 13 model-and-domain cells and tied in the last — 28 points worse on average. Three quarters of that loss is conversations that ran out of budget before they finished. Remove the ceiling and the split draws level, at 5.5× the tokens.