Is my agent workflow getting cheaper? A dashboard that reports an honest null
Cost per point from my own session logs, sizes from a frozen judge, parts that refresh every week.
Built 2026-10-05 · Filmed 2026-11-05
I keep a page that asks one question about my own work: is each unit of finished work with AI agents getting cheaper or dearer? Its latest reading says the cost per unit fell by 22 percent, and its own range for that change runs from a 70 percent fall to a 105 percent rise. A range that crosses zero cannot tell a fall from a rise, so I do not claim that cost fell.
What you're watching
The page is one long, quiet dashboard. It opens on its question, then gives the headline: what a point of delivered work cost over the last 28 days, with a line next to it saying the level is uncertain by about two to three times. Beside that sits the change over the last 14 days, drawn as a dot on a bar. The bar is the 90 percent range, and it crosses zero.
The tour then follows the page down. The ten factors that add up exactly to the change. A monitor of five rows, checked every week, that cannot raise an alarm yet. The cost per point day by day. Two parts that move every week: output and cost per turn at each effort level, and where the money sat. It ends on a list the page itself calls "What to distrust".
Three plain words carry the page. Cost is token cost, read from my own session logs and priced at one constant price book. A point is a size score for a piece of delivered work. The size comes from a frozen judge, and a weekly timer rebuilds the parts that move.
Nothing on it is argued on the day: the numbers are computed. Newer work is sized by a frozen judge: the same model, the same reference projects and the same prompt, three placements, the median kept. The change gets its range by redrawing every project's size 1,000 times. Where the data cannot carry a claim, the page shows it: the range crosses zero, the page calls the level uncertain by about two to three times, it says no row of the monitor can raise an alarm before a fixed date, and it lists a check that should have caught a real effect and failed. The point of this card is that the instrument exists, is computed by fixed code from logged data, and reports a null honestly. The one model in it is the frozen judge, and the page measures that judge's noise. The card does not say cost fell.
What doesn't come across on video: the page also has parts on how much room each group of factors has, on my own attention in hours and a glossary, and the film skips them. The table that lists every project is left out of this copy, so no project is named on screen. The figures are the ones the page showed on the day of filming, not a live view.