Every game on our games page was written by AI agents, mostly Claude and Codex. We describe what we want, an agent writes the code, then we play it and say what feels wrong. This post is about the end of that loop: how a game gets from an agent's folder onto the site, and why some of the cards say Beta.
What the agents do
An agent builds the game: the rules, the screens, the sound, the offline cache and the files that let you install it on a phone. It writes tests for the rules and runs them before every build. It also gets a game ready for the site. It fixes whatever our review script flags, writes the text for the card and makes the picture people see when a link is shared.
As we went, we wrote down what worked in each game as recipes the agents read before starting the next one: how the screen should sit on a phone held sideways, how an update should wait until the player taps it, how a game keeps playing with no signal. Seven games have gone into those notes so far.
The checks a script can do
Since 19 October every game has to pass our store standard before it gets a card. A script serves the build the way the site will, opens it in Chrome as if on a phone and checks that it:
- opens without errors or missing files
- keeps working with the network off
- plays sound from the first tap and has background music
- saves under its own name, so it cannot overwrite another game's progress
- fits four phone sizes without sideways scrolling
- shows a proper title, description and picture when the link is shared
It also pretends a new version has been released and checks that the game tells the player, without reloading in the middle of a round.
Most of these checks exist because a person noticed something. Hội Báo Xinh was the first game through the process, and its first upload passed only the install checks. Mas played it and said it was missing sound, and that a slot game needs its big-win moments. Then a shared link showed no picture. Each miss became a rule and each rule became a check, so the next game cannot slip through the same way.
What a person still has to do
Some checks cannot be handed to an agent, and the standard says so: an agent may not tick them.
- Play it on a real phone. Install it, open it from the home screen, play for ten minutes, then switch the network off and keep playing. The script only imitates screen sizes.
- Play it through. From the first screen to a real ending, written down: what we played, what we saw, what we fixed.
- Say whether it feels right. A simulation can show the first chapter is winnable. Whether it is too easy or a chore is for someone playing it to say. Sound is the same: the script can tell there is music, and a person decides whether it is any good.
- Look at it. The script saves a screenshot at every phone size. It finds a layout that is broken; whether a layout looks good is a person's judgement.
- Decide. Which tier a game goes up at, what the card says and when it is published are a person's call. Agents publish only when told to.
We had a small example this week. Asked to "remake the thumbnail" for Buddy's Number Nook, Claude remade the square picture on the games page. Mas meant the picture shown when the link is shared. Nothing broke and no check would ever have caught it. It took a person saying "I mean the one when sharing".
Why a game says Beta
A game goes up at one of two tiers.
| Tier | What it has passed |
|---|---|
| Beta | Every automatic check, plus a person has played one full round, checked the sound and played it on a real phone |
| Released | All of that, plus a recorded playthrough to the end, a balance report with its targets set before the numbers came in, licences for any third-party art, music or fonts, and a version number in the game |
A Beta game is playable and installable, and it works offline. It has not yet been played to the end by a person who wrote down what they found.
Buddy's Number Nook went up on 20 October and says Beta. It passed the automatic review the same day. The real-phone test, the playthrough and the balance report are still to do, and the sound settings have on and off switches but no volume control yet. Strictly, our own standard asks for the phone test before Beta, so this one went up a step early. Hội Báo Xinh is the other game marked Beta.
This post was drafted by Claude Opus 5.5 on 20 October 2026, in the session that put Buddy's Number Nook on the site.