I Have a Day Job and Built 3 Apps in 2 Months. AI Agents Did All the Work; I Only Made Decisions

On June 28, 2026, I gave my first AI agent the role of “PM.”

Over the next 2 months, 3 apps came out of it. One is live on Google Play, one has passed review and is waiting to go live, and one is in closed testing.

I have a day job during the week. I’ve written almost no code. All I did was make decisions.

This site is where I keep the record of that.

What I built

Kiroku — a one-line AI diary

Kiroku's store listing image. Mood colors line a calendar, and a graph of the month's feelings is shown — Each dot on the calendar is that day's mood color. The longer you keep at it, the more you see your own ups and downs

It’s a diary app where you write just one line and the AI replies with a comment. It scores your mood from 0 to 100 based on what you wrote, and the calendar gradually fills in with mood colors.

I built it on the hypothesis that “maybe people can’t keep a diary because there’s too much to write.” Of the 3, it’s the only one already live on Google Play.

Nemuru — AI bedtime picture books

Nemuru's story playback screen. The text of the generated picture book is shown, with a read-aloud speed toggle and a play button at the bottom — A different story is generated every night. You can switch the read-aloud speed to "slow" because that turned out to be genuinely necessary for bedtime

Enter your child’s name and what happened today, and it generates a picture book just for that child and reads it aloud. It’s passed review and is waiting to go live.

Narabu — AI queue wait-time prediction

It’s an app that predicts the wait at popular restaurants with lines before you get in line. It shows them on a map color-coded green, yellow, and red. When people who actually lined up report their wait time, those real measurements correct the predictions every day.

It’s in closed testing right now, and I’m wrestling with Google Play’s “12 testers × 14 days” requirement.

The setup

I run a separate AI agent session for each role.

                        Human (decisions only)
                                   │
                                  PM
       ┌─────────────┬─────────────┼─────────────┬─────────────┐
Implementation     Legal        Design     Monetization     Growth

Each app has this structure, so around 30 “employees” are running at any given time.

My role has become nothing more than answering the decisions that get escalated to me. The amount of work went down, but the density of decisions clearly went up. If you start out expecting “AI will make this easy,” you’ll probably be disappointed.

What I’ll write about here

There are already plenty of how-to posts on “building AI employees.” Adding more of the same would be pointless.

What I’ll write here is a record of what I actually did, what actually broke, and what it actually cost.

Three posts are already up:

My cloud bill for a month was ¥22, and I actually paid ¥0 — with real data from the invoice
I renamed my GitHub account and the policy pages for all 3 apps went 404 — the full record of an incident that left me unable to submit for review
3 things that worked with 30 AI employees, and 3 that were a complete waste — why starting from an org chart was a mistake

What I plan to write next:

How many times I’ve failed Google Play review (ongoing)
How I rounded up “12 testers × 14 days” as an individual
The actual revenue from all 3 apps (this won’t be a simple “¥0” story)
How much I pay my AI employees each month

Writing only about what went well probably wouldn’t help anyone. I’m going to write about the failures in detail.

The apps I’ve built are collected under Works. Updates go out on @YKStudioLab on X (Japanese).