Week 24: the machine starts selling.

The three-week silence

Every agent call had been failing since late May — the free-tier model route had a token-per-minute cap smaller than a single agent prompt, so every request bounced with the same error. Processes ran, schedules fired, dashboards looked alive. Nothing thought. The fix took a day: reroute every agent to a larger free route with a local model as backup, then add a canary that tests outcomes every 30 minutes instead of trusting that running processes mean working agents.

What shipped after that

• The store went live. Five original designs — anime desk mats and a tarot poster — generated locally, quality-checked by a local vision model, and published on demand. See the shop.
• Product walkthroughs in seven languages. Free and Plus, narrated and subtitled in English, Swedish, Spanish, German, French, Italian, and Portuguese — all rendered locally.
• The GPU got a night shift. One 16 GB card shared by video render, image generation, and QA models was capped to a 4-hour daily window. Now it works 21:00–11:00 and drains the queue while the studio sleeps.
• Quality gates that say no. A local vision model rejected two renders for visual artifacts before they reached the channel, and the music gate is currently rejecting an entire batch of ambient tracks until the generator improves. Annoying. Correct.
• The org talks to itself. Daily agent standups land in a shared hub, a morning report summarizes the last 24 hours, and a daily self-verification checks that the whole overnight cycle actually ran — and says so, honestly, either way.

What didn't work

Plenty. A retry loop re-rendered the same failed batch into 85 GB of disk garbage before a cap was added. The first three published products were pulled the same day for quality. A segment library starved itself because publishing permanently banned clips faster than rendering made new ones — fixed with a 45-day reuse window. Each failure is now a rule the system enforces by itself.