Demos are free. Systems aren't. A thing that runs every day has a metabolism, and most of what I've learned this year is about feeding it without waste.
The model calls are the obvious line item, and the cheapest lever there isn't a discount, it's not making the call at all. A lot of work doesn't need a frontier model. Routing the cheap, mechanical steps to a small local model and saving the expensive reasoning for the moments that actually need it cut my spend more than any pricing change ever did.
The less obvious cost is electricity and heat. The orchestration host is a local GPU workstation, and it runs warm. I watch thermals the way you'd watch a stove you left on. A system that runs unattended overnight has to be one you trust not to cook itself, which means monitoring is not optional, it's part of the build.
And the cost nobody warns you about: attention. An autonomous system generates a steady stream of things that look like they need you and mostly don't. The real skill is building it so it only interrupts you for the decisions that are genuinely yours, and handles the rest on its own. Most of my safety work is really about earning the right to not be in the loop.
← More field notes · Want to talk it through? Get in touch →