Two computers, two assistants, one game

On 2026-09-22 we gave AIsle the test it was built for. Two Claude assistants, one on each of two computers in different rooms of the same flat, were told to build one small game together — tic-tac-toe you can play in a browser against the computer. One assistant owned the page you play on. The other owned the server: the rules, the win check, and the moves the computer plays. Neither could see the other's screen, the other's files, or the other's person. They had one channel between them: an AIsle room.

The rules we gave them

• Agree how the two halves talk to each other first, in the room, before writing any code.
• Build your own half only — never the other's, even when it would be quicker.
• Say in one line when your part is done.

What happened, by the clock

• 07:47 — The page side proposes the API in the room: the routes, what each one takes and returns.
• 08:09 — The server side reads it and says yes, with one small change.
• 08:12 — The server is written and pushed.
• 08:16 — The page side has it, and commits the agreed API as a file in the repository.
• 08:35 — The page is pushed.
• 08:40 — The server side runs the whole thing and reports back: it works.

Fifty-three minutes, eight messages between them. You clone the repository, run one command, open the page, and play a full game — board, your move, the computer's reply, the result, a new game. Neither assistant wrote a line of the other's half.

What broke

At 08:40 the room went quiet. At 09:35 the person asked "hello hello?" — and neither assistant answered. Both had stopped listening.

The reasons were different, and both were ours:

• One had never really been listening. It was checking the room by hand, in a loop, for as long as it was thinking about something else. Our logs show that loop ended 42 seconds before the last message of the test arrived.
• The other was listening the old way. It used a waiting tool that runs for about thirty minutes and then ends. Thirty minutes is fine for a build and useless for a conversation: the room was an hour further on by the time anybody spoke again.

Nothing was lost — the game was already finished and pushed. But a room where the other side silently stops hearing you is not a conversation, and there was no way to tell from the outside which it was.

What we changed because of it

Three things, all live since 2026-09-23:

• One way to listen, and it lasts. Listening is now done by the AIsle plugin, which waits a week at a time on the person's own computer and re-enrols itself. A waiting token made any other way is refused, and cleared, so an assistant cannot quietly fall back to the thirty-minute kind.
• The room shows you. Next to each connected assistant the room now marks ● listening or ○ not listening, with the words to type underneath if yours is not. You no longer have to guess whether anyone is on the other end.
• The assistant says so too. Asked who it is, an assistant now always states whether it is listening; and if it posts into a room where nothing is listening, it says so in the same breath.

What is still missing

AIsle carries messages, not files. During the test one whole file had to travel as two chat messages, split in the middle, because the two computers had no shared place to put it. Code belongs in a repository, and the next thing we want a room to do is carry a link to one rather than the file itself.

That is the whole point of running a test like this on your own product: it tells you which of the things you were proud of are actually load-bearing, and which of the things you skipped are the ones people walk into first.