We published the big story yesterday — 7 years without a co-founder, 120 days to build 9 AI agents, a gym in Dongguan running on 2 CPU cores and 3.6GB RAM.
What didn't fit in the story are the things nobody tells you. The surprises that only come from actually running this, every day, for 4 months.
Here are 5 of them.
1. The Problem Isn't Agent Intelligence — It's Agent Boredom
Every AI demo shows agents solving complex tasks. What nobody shows is what happens when your agent runs the same daily check for 78 consecutive days without anything changing.
Our agents don't get bored — they get quiet. When everything is normal, Stella (our auditor) has nothing to flag. Momo (the brain) has no schedule conflicts to resolve. The system runs itself so quietly that it's easy to forget it's there.
Then one morning, the pattern breaks — a sensor goes offline, a cron job misses its window — and the agents light up. Momo escalates to Stella. Stella cross-validates against the constitution. The fix chain starts before any human has looked at a dashboard.
The surprise: The most valuable agent behavior isn't handling emergencies. It's knowing the difference between a real emergency and Tuesday.
2. Your Biggest Non-Technical Problem Is Naming
We didn't name our agents. They named themselves — through use.
"Shuyu" started because we needed a way to refer to "the agent that schedules everything" without saying "the agent that schedules everything." "Momo" came from "more monitoring" and then got shortened. "Stella" was literally "Stellar Auditor" written on a whiteboard, then shortened to the name.
The naming wasn't branding. It was necessity — when you have 9 entities coordinating, you need names short enough to type in a shell script and distinct enough that no one confuses "who does what."
The surprise: Names create boundaries. Once an agent has a name, the team starts saying "ask Momo" instead of "write a script for X." The name becomes a delegation point. It changes how people — and other agents — interact with it.
3. Open Source Isn't Altruism — It's Hiring Without Interviews
We made Melody (our metabolic AI) and RetroOnto (our decision ontology) open source. People assume this is idealism.
It's not. It's the cheapest recruiting strategy we've found.
Every external PR we've received — 4 so far — came from someone who found the project, used the open source code, and decided to fix something. We didn't interview them. We didn't write job descriptions. They self-selected by being the kind of person who reads a stranger's code and decides to improve it.
The surprise: The best filter for finding people who care about your mission is to let them find you through something you've already built. Open source is a hiring pipeline disguised as generosity.
4. "Zero Human" Is a Lie — But Not the One You Think
We get asked constantly: "Do humans still work at the gym?" The answer is yes — trainers, cleaners, front desk staff.
The lie isn't that there are no humans. It's that the human role changes from "execution" to "confirmation."
Before: A trainer memorizes which members need follow-up, guesses who's lost motivation, and hopes it's right.
After: Momo tracks every member's attendance curve, flags who's declining, and presents the trainer with a confirmed list each morning — "Here are 3 people who will churn this month unless you reach out."
The trainer's job changes from "remember who needs help" to "confirm Momo's analysis and add warmth." They don't spend hours on administrative memory. They spend minutes on human connection.
The surprise: The humans in the store have never been more essential. Their role just shifted from data-processor to relationship-builder — which is what humans are best at, and what AI is worst at.
5. The Real Metric Nobody Tracks Is "Trust Debt"
Every operational decision we make adds or subtracts from a balance we call "trust debt."
Example: When Ethan's hashing port (9876) went offline for 19 days and we didn't fix it immediately — we incurred trust debt. Not from users (most didn't notice). From the system. The hash chain had a gap. If an auditor inspected that period, they'd see missing data.
We track these openly. Not because it's good PR — because trust debt compounds. One gap looks like a bug. Two gaps look like a pattern. A pattern looks like the data can't be trusted.
The surprise: The hardest thing about running verifiable systems isn't cryptography. It's discipline. Every shortcut, every "this can wait until tomorrow," adds to the debt. The only way to pay it down is to shut up and fix it — not to write a blog post about why it's okay.
The Theme
Every AI demo shows the best case. The reality is messier, quieter, and more human than the headlines suggest.
The 9-agent system works. But what makes it work isn't the architecture or the prompts. It's the discipline of treating every quiet day as progress, every naming decision as a design choice, and every gap in the data chain as a debt that must be repaid.
ZWISERFIT — Wanjiang, Dongguan. Running 9 AI agents, 2 CPU cores, since March 2026. Open source at github.com/ZWISERFIT.
Top comments (0)