The demo is my house.

One property — house, greenhouse, lawn, theatre, print farm — run by a fleet of 38 scheduled AI agents on top of Home Assistant. Morning standups, inbox triage, greenhouse and lawn watchdogs, CEO-style desk reports for each small venture, a marketplace scout, and a night audit that re-reads the day's work.

The agents aren't the interesting part. The instrumentation is. A run recorder logs what each agent actually wrote. A watchdog compares file timestamps against the scheduler, because “the task fired” and “the work happened” are different claims. The night audit uses a second, independent model to cross-review — one model never grades its own homework.

That design earns its keep. A single audit pass caught four agents that had been firing on schedule and writing nothing — one of them quiet for two weeks. No log line would have told you that. A file timestamp did.

If an agent can fail silently, it isn't finished. That rule is most of the product.

TURF is a micro soil-moisture sensor that installs like an irrigation peg — a Ø64 mm cap carrying a 50×50 solar cell on a Ø25 shaft, sitting 10 mm proud so the mower passes straight over it. It goes into the ground with a printed strike puck instead of a shovel.

It went from rev B to rev G in a day: parametric CAD, mesh-verified STL, firmware, and live readings in Home Assistant the same evening. Not heroics — a short pipeline. When design, verification, printing, and flashing share one bench, an iteration costs an hour, so you spend geometry instead of meetings.

The revisions were earned, not cosmetic: rev A trapped its own battery (component fit is now checked against the shell before anything prints); the head-house got an o-ring seat and a proper wire path; rev G swaps radio for a wired RS-485 Modbus variant with a parametric probe socket — drawn, verified, and deliberately unprinted until the real probe is measured.

An iteration that costs an hour changes what you're willing to fix.

putahatonit.ca sells 3D-printed hats for robot mowers and vacuums. It also happens to be the testbed for a full AI product pipeline: design → render → copy → staged social, with a human approval gate on everything that ships or publishes.

The honesty rules are mechanical, not aspirational. Product renders are built in Blender from the actual print CAD, in the actual filament colours pulled from inventory. AI lifestyle images are allowed — labeled as impressions — and must pass a QC gate: deterministic checks, then a vision model, then a person. The gate has already bounced an AI image that forgot to include the product.

The catalog rebuilds automatically from approved SKUs only; the site is static and framework-free on Cloudflare Pages, like this one.

putahatonit.ca →