Or how to produce a €60 lettuce. Eleven sensors, a camera, an old phone and an agent that waters my vegetables.
Read in full here:
Funny that the €300 of sensors is the big line item and the agent reading four million rows every morning comes out at “cents”. That holds while the agent queries the tables and gets back a small summary; once it starts pulling the camera photos into context every day, the bill moves fast. At aicost.tools (I edit it) we keep an agent cost calculator for exactly this kind of back-of-envelope check: AI Agent Cost Calculator — What One Agent Task Really Costs · AI//COST Does the daily run look at the photos, or only at the sensor tables?
I’d keep the daily run off the camera unless the agent actually needs visual state. If the camera is mainly checking for obvious problems, run a small local classifier or simple image diff first and write a short status into the lakehouse; let the agent query that summary with the sensor rows. Full images can then be pulled in only when something is ambiguous or a sensor trend looks wrong. That keeps the cheap path cheap and makes the occasional expensive inspection intentional.