A conversation about an agent's cost almost always veers toward tokens: how to count them, where to trim, which cheaper model to take. The instinct is understandable but misleading - it optimizes the unit of spending instead of its cause. Tokens are not spent on their own; they are spent by iterations, and the number of iterations is set by how much unresolved uncertainty is left in the task.

The naive move is to save on the input. Take a weaker model, trim the context, run the check less often - on paper each of these actions lowers the bill for a single turn. In practice it more often raises their number: understood worse, did worse, asked again, redid. Saving on a turn turns into extra turns, and the total spend grows exactly where you tried to cut it.

The naive economy breaks on the fact that the main driver of cost is not the price of a turn but the count of turns, and that is driven by uncertainty. Spending grows from wide context, from repeated investigation of the same thing, from a strong model where a simple one would do, from subagents and parallel variants launched without need. All these line items share one thing: they multiply work that was not narrowed in advance. It is expensive not to think but to guess many times over. The practical takeaway is one: the bill grows not so much where an expensive model is chosen as where the task is posed so that it must be walked through twice and three times.

Hence the reversal of strategy: what should be cut is not verification but uncertainty. A small task instead of a vague one, a precise reproduction instead of "it fails sometimes", a saved plan instead of a spoken one, narrow commands instead of walking the whole repository - each of these techniques cuts not the cost of a single step but the very number of steps. Verification stays in place: to cut it is to save on the one thing that tells the done from the claimed.

Before an expensive turn it helps to ask it one question: which unknown does it remove. If there is no answer, the turn investigates at random, and its cost almost certainly returns as a second pass. If there is an answer, you are paying for a specific removal of uncertainty, and the result of that turn is worth saving. The question "what will I learn" turns spending from a bet into a deliberate purchase of knowledge. And the reverse: a turn with no clear answer to that question is most often a signal to step back and narrow the task, not to launch an investigation in the hope that it stumbles onto the needed thing on its own.