Main idea

A skill has to analyze three customer records. For lack of a filter at intake, it loads five hundred. The result produced is identical; the spending is not — roughly 150,000 input tokens for material that will never serve.

The cost of a read doesn't follow what you get out of it, it follows what you put into the window. An agent doesn't "skim": everything that comes in is paid for, including what it discards immediately.

Hence a shift in the optimization work: the question is no longer how to ask better, but how to query the database — filter, pagination, prior selection by script. Conditional reading goes further still: not opening a source at all when a prior condition isn't met.

Why it matters

This places the economy on the data-access side rather than the prompt side, where it is usually looked for.

And it gives a symptom you can spot in an audit report: a read step whose volume bears no relation to the number of items actually processed.

Nuances and limits

Filtering presupposes knowing in advance what is relevant. When relevance is precisely what the run has to establish, reading broadly isn't waste, it's the work.

And a badly framed selection is worse than a complete read: it produces a plausible answer on truncated material, with nothing to flag it.

Open questions

• How do you measure, after the fact, the share of loaded data that actually weighed on the output?