Every morning I open one page before I open anything else. I call it HQ. It’s the home page of the system that runs my vape stores, Divine Tribe, Nice Dreamz wholesale and Tribe Seed Bank, and its whole job is to answer two questions fast. What did the agents do for me while I wasn’t looking, and is anything broken right now.
This morning it said 57 agents, 6,310 runs today, 40 pieces of real work, and nothing broken right now. That last line is the one I actually read.
At the top are the things that need me. Orders and shipping at the shop on one side, customer questions that need an answer on the other. Below that is the part this post is about, a feed called “what your agents did.” Every agent that actually did something today gets a card with one plain sentence. Failed orders rescued an order after the card was declined. Inbox reaper went through 32 unread emails and binned the one that was junk. GitHub backup backed up all 56 repos and 2 gists in four minutes. Toll booth counted the AI bots that visited my sites yesterday.
Agents that only checked and found nothing don’t get a card. They roll up into a strip of small chips at the bottom called quiet watchers, with a count of how many times each one looked. Hostinger guard checked 392 times this morning. I don’t need to see 392 cards that say “all good.”
Until yesterday, the feed had a problem. Eighteen of the 31 cards were blank. Some agents had little clips I’d rendered back in September, and the newer ones had nothing behind their text but a dark gradient. The page told me what happened but it didn’t look like anything, and the cards all blurred together.
So I gave every agent its own scene. The rule was simple. You should be able to tell what an agent does from the picture alone, before you read a word. Toll booth is a little robot in a striped booth with a clicker counter, waving through a duck car, a flower car and a search bot car. Failed orders is a kid in a red cape diving across a shop counter to catch a credit card before it hits the floor. New model watch is a lookout in a crow’s nest with a brass telescope, spotting a glowing ship coming in over the clouds.
Before making 34 of them I ran a style test. Three agents, two looks, a painted storybook style and a glossy 3D movie style. I went with painted.
Everything was made on my own machines with local models, with nothing paid and nothing sent to a cloud service. The stills came from Krea 2 on my Mac mini at about two and a half minutes each. I looked at every one and redid ten that had garbled lettering or put the action where the card’s text would cover it. Then my M5 animated each still with LTX 2.5 at about 104 seconds a clip. All 34 rendered between 1:46 and 3:15 in the morning with no failures. Each card plays a four second loop, two seconds forward and two seconds back, so there’s no seam where the end jumps to the start.
While I was in there I fixed the words too. A few cards were still showing raw machine output like “done rc=0” and “exit 1: OK-held.” Now every card reads like a sentence a person would say out loud, using the agent’s own numbers. And one card I like a lot is called Fixed itself. When something breaks and then heals on its own, it doesn’t stay red on my screen. It folds into that one green card, so I’m only warned about problems that are still happening.
At the very bottom is a status bar with how many sites are up, how many unread emails are waiting and how many agent runs have happened today. Seven of seven sites were up this morning.
HQ itself is private, because it touches orders and customer email. But I wrote up how the whole system works, from checkout to the label to the support replies, and that’s public here: https://github.com/nicedreamzapp/how-my-shop-runs
The paintings are a small thing. But I open this page every day, and now I actually enjoy looking at it. When a new agent starts working, it gets its own painting the same night.