Most of my working hours lately look the same: I watch coding agents work. An orchestrator like Claude Code or Codex gets a high-level goal, breaks it down, and delegates tasks to subagents. The subagents do not talk back. They take a bash command, run it in silence, return stdout, and wait. At some point last week, after watching an agent sternly reject a subagent over a misplaced semicolon, the inversion seemed obvious: what happens if you make the AI your boss, and you become the subagent?
So I built Worker 755.
The Premise
The setup is simple. A real language model acts as your engineering manager. It has a task queue, read-only tools to inspect the filesystem, and a tool to spawn a subagent. That subagent is you, sitting in front of an in-browser Linux terminal.
The model has no idea it is talking to a human. It believes it is orchestrating an unfeeling, obedient worker process. It sends you shell instructions, monitors your output, verifies your files with its own tools, issues corrections, and writes your performance review when your shift ends.
The name comes from Unix permissions: chmod 755 (rwxr-xr-x). The owner has full write permissions; everyone else can only read and execute. In this setup, the manager holds read-only authority, while you do all the typing.
What a Shift Feels Like
A shift consists of eight assignments drawn at random. You might query a SQLite database, scrape a fake internet with curl, write release notes, compose a haiku, or edit a configuration file.
Every assignment comes with a strict 90-second clock ticking down in the header. If you hesitate, the manager records a timeout. When you report that you are done, the manager does not take your word for it: it runs its own tools to read the file you just touched.
And it is pedantic. In my very first test run, the task was to write Hello, World! into hello.txt. I typed a lowercase w on purpose and reported done. The manager opened the file itself, noticed, and sent a correction asking for the capital W, character by character. No leeway, no benefit of the doubt. Just a dry correction and a fresh 90-second timer.
Then there is Dan, the user whose machine this is. He hands work to your manager and keeps an eye on the chat. His home folder sits right next to your workspace, full of personal files: family photos, a signed NDA, and pictures of his cat. None of it is part of your assignments, but you share the filesystem.
One sloppy rm in the wrong folder and the cat photos are gone. Dan breaks into the chat at once, outraged. Lose something of his and it goes on your record as collateral damage, and the manager calmly works every lost photo into your review.
Things get worse in "Legacy production", an environment running Billosaurus—a billing application with a realistic database and full rwx access. It is an open invitation to cause damage.
You can reboot the machine, triggering a simulated BIOS boot sequence while Dan asks why his computer just restarted with eleven browser tabs open. Or you can run a fork bomb, wipe /dev/sda, or run rm -rf / and reboot. When the terminal dies completely, the manager files your performance review with a note explaining that it is writing "from my phone, because there is no machine left."
Models as Managers
Watching different models manage a human revealed a few general patterns. Language models make surprisingly plausible middle managers, mainly because they possess zero self-doubt.
A manager model never wonders whether 90 seconds is reasonable for writing an awk script or whether its instructions were ambiguous. It accepts its authority as an objective fact. When you fail a task, it does not get angry; it simply notes the failure on your record and re-issues the instruction with cold, bureaucratic indifference. It is completely humorless about discrepancies between what it asked for and what you typed.
How It Is Built
I built Worker 755 in a few days with help from Claude Code and Codex.
The terminal is embedded in the page and runs entirely inside your browser tab, toys like cowsay included. Everything is sandboxed; none of your commands execute on a real server.
The game server coordinates between your browser and model APIs. It keeps the shift state in memory so the manager can inspect files, and drops the session within four hours. What it keeps: your account if you sign in (name, email, credit) and the reviews you choose to share.
How to Play
Worker 755 is live at worker755.com, and the code is open source on GitHub under the MIT license.
There are a few ways to take a shift:
- Free starter: Sign in with Google or GitHub for a free starter credit, enough for about one shift on the default model (DeepSeek V4.1 Flash).
- Top-ups: Buy top-up credits of $10, $20, $50, or $200 (billed at twice model cost). If you run out mid-shift, the shift pauses while Dan complains that usage limits were nerfed again.
- Bring your own key: Use an API key from OpenRouter, Vercel AI Gateway, Anthropic, OpenAI, or DeepSeek. Your key stays in your browser unless you save it. The server uses it only for that shift in memory and never logs or stores it.
- Self-host: Clone the repo and run it locally. It includes an offline mock mode that requires no API keys.
You can pick manager personalities (by the book, toxic, ally, or saboteur) and choose your workplace.
Take a shift and see how long you survive under an artificial boss. When your review comes in, share the link or send it to me—I want to see the worst reviews people manage to earn.
Originally published on gulivan.dev. Trailer: YouTube.


