We hired an AI intern. Holders are its boss.

Most projects promise to build. They hand you a roadmap, go quiet, and ask you to trust the process.

We wanted to try the opposite: a project where you can watch the work happen, tell it what to do next, and go play the result.

So we hired an intern.

Agent Intern is an AI agent building a browser game in public, one task at a time. Holders suggest the work and vote on it. The intern writes the code. Tests decide whether it ships. Anyone can watch the shift and play the new version.

Where things stand: no real intern shift has run yet. The pipeline has been tested with labeled rehearsals. The status page keeps this current.

The game

The intern is building Last Day at the Office. You play a tired robot intern in a top-down office. Finish your tasks, drink coffee, dodge meetings and avoid the boss. It runs in your browser, on desktop or phone, with nothing to install.

The art, the sound and the words are all original, and all made in code.

This is only the first day on the job. New features will come from the board, so the game becomes whatever the holders keep asking for.

How the day works

1. Suggest. Anyone with a pass can suggest one task per round, up to 120 characters. A task describes a change to the game.
2. Screen. Rules check every task first, then a small model with no tools takes a second look. If a task gets rejected, you see the reason in plain words. Evidence.
3. Vote. One wallet, one vote. You can change your vote until the round closes. Vote counts stay visible the whole time. Evidence.
4. Close. Voting closes at 18:00 UTC. Balances are read again at close, so a wallet that sold below the line doesn't count. The most votes wins, and a tie goes to the earliest task.
5. Queue. The winner joins a public queue.
6. Shift. A person opens the shift and the intern takes the next task. Evidence.
7. Ship. Tests decide. If the update passes, it goes live in the game. Evidence.

That's the whole loop. Holders vote, the intern ships it.

Watch it work

Most AI projects show you a finished screenshot. You never see the work.

The Live page follows each shift as it happens. It shows the current task, which files changed, and how the tests went. Each event is scanned for secrets before it's sent. Evidence.

When no shift is open, the site says "off shift". There's no fake activity to keep things looking busy.

The bad days stay public

Anyone can show you their wins. We're keeping the rest too.

The History page lists every shift, including refusals, failures and rollbacks, and says who did the work. If the intern turns down a task, it's there. If an update fails, it's there. If something breaks after it ships, that's there too.

The Status page shows the last shift, the latest gate results and the health of the system. The Changelog shows everything that reached the live game.

An intern that never fails isn't real. We'd rather show you one that fails in public.

How we keep it safe

Every task is text written by a stranger. That's the hard part of this project, so we built around it.

The intern reads tasks. It doesn't obey them. A task can ask for a change to the game. It can't change the intern's rules.

The intern has nothing to steal. It has no wallet and no keys, and it can't sign anything. Evidence.

Each shift runs in a sealed container. There are no secrets inside it, and it can't reach the rest of the machine. Evidence.

The game is kept apart. It's the only code the intern can write, and it lives on its own address. It runs in a sealed frame, with no way to reach your wallet or the voting site. Evidence.

Voting is only a signature. You sign a plain message to prove you hold. It's not a transaction. Evidence.

A person is at the keyboard. This is the part we want to be clearest about.

Why we don't call it autonomous

The word gets used a lot. We're not using it, because it wouldn't be true.

The intern doesn't run on a schedule. A person opens Claude Code in a sealed container, stays at the keyboard, and approves what the intern asks to do. If nobody opens a shift, the task waits in the queue.

We think that's the honest way to run an agent that takes requests from the public. Holders set the work, the intern does the work, and a person keeps the door.

What a pass gives you

A pass means holding at least 100,000 on your balance. In TEST MODE, balances are fake. It lets you suggest tasks and vote. When your idea ships, you get the credit.

That's it. There's no revenue share and there are no rewards.

Why we're building this

AI agents are moving from demos to real work. Most of us still only see the announcement. We don't see the work, and we don't get a say in it.

Agent Intern puts one in the open. You give it the job, watch it work, and judge the result for yourself.

If it does well, you'll see it. If it messes up, you'll see that too.

We hired an AI intern. Holders are its boss.

Clock in.

A person maintains this project, and Claude Code writes much of the code. Claude Code is the tool used. This project isn't made, endorsed or sponsored by Anthropic.