I gave my AI assistant a self-check. Every part gets a light, from Online to Down, plus the one step that fixes it.
On the live data, it would already flag two misses: no morning brief on Wed Nov 4, and a weekly review that has never filed.
A missed run throws no error.
5/
What helps isn't a smarter model. It's a board.
The board I built makes an agent claim its work. A taken migration number is refused. Look-alike work has to say how it differs. Claims lapse after 8 idle hours.
How do you keep parallel agents from stepping on each other?
1/
40 pull requests merged into my personal AI assistant's code between Friday and Sunday night. AI coding agents wrote them, several at a time.
Parallel agents break things a single agent never does. Here's what I ran into:
4/
Stale rebases. One feature was rebased before a sibling landed, so its list of read-only tools was missing a new one.
In a practice run, a harmless lookup would have counted as an attempted action. One-line fix, the next morning.
About 430 name ideas for a new project in one day.
Most good words had the .com taken. The one I loved belonged to a US company in the same field, registered last year. Its domain was free, so I almost missed it.
Now I search name plus industry first. What did your search miss?
My app never asked for the microphone. A library did.
An audio library quietly added the permission to the final build. It would have shown up on the store listing.
Check the final permission list, not just your own code.
5/
The second pass mattered as much as the first. It withdrew one bug that wasn't real and corrected another.
Before testers now, I test on the devices and sites people actually use, and a second pass checks the first.
What's the smallest screen you test on?
1/
Before I handed a side project to testers, my AI coding agents ran a full QA pass. Then a second pass re-checked every finding from the first.
What they caught:
Running guest support taught me not to trust one number alone.
I watch three together: cost per interaction, guest satisfaction and QA.
A team can be cheap and still wrong, or polite and still sloppy.
Which number would you watch first?
A bug from my family-story app: on playback, browser-recorded videos didn't report their length, so the player read the duration as Infinity.
Fix: save each recording's length when it's made, and have the player use that.
Don't trust the browser about media you recorded yourself.
Bug on my family-story app: on phones, the dashboard scrolled sideways.
The page was 720px wide on a 390px screen. The main column couldn't shrink and the menu didn't wrap.
Fix: let the column shrink and make the menu scroll inside the screen. Now it's checked automatically at 360, 390 and 430px.
I run guest support for US vacation rentals from Venezuela. Midmost, 8 years.
I build software with AI and share what works in ops and in code: a family-story app in private alpha, plus AI assistants that work like extra staff.
Misses included. DMs open if you run a service business and build with AI.