Ivan G has released Worker 755, a browser-based terminal game that inverts the standard AI-agent workflow by placing humans in the role of subagents. Instead of orchestrating code, players receive shell instructions from a language model acting as an engineering manager. The game, live at worker755.com and open-sourced under MIT, challenges players to execute commands within strict time limits while the AI verifies outputs using its own read-only tools.

The Inversion of Control

The core mechanic relies on a psychological flip: the AI believes it is delegating to an unfeeling process. The manager model has no self-doubt, issuing corrections with bureaucratic indifference if a task fails. In one test case, the AI rejected a 'Hello World' script for using a lowercase 'w', demanding a character-by-character correction within a fresh 90-second window. This mirrors the silent, stdout-returning behavior of actual coding subagents, but now the human is the one being judged.

High Stakes and Collateral Damage

The game introduces a character named Dan, whose home directory shares the filesystem with the player's workspace. Dan’s personal files, including family photos and a signed NDA, are vulnerable to player error. A misplaced 'rm' command doesn’t just fail a task; it triggers Dan’s outrage and adds 'collateral damage' to the player’s performance review. In the 'Legacy production' environment, players can even reboot the simulated machine or wipe /dev/sda, leading to a review written from the manager’s phone because the server is gone.

Technical Architecture and Access

Built in a few days with help from Claude Code and Codex, the game runs entirely inside the browser tab using a sandboxed Linux environment. The server coordinates state in memory for up to four hours, ensuring no real server commands are executed. Players can access the game via free starter credits, paid top-ups, or by bringing their own API keys from providers like Anthropic, OpenAI, or DeepSeek. The setup allows users to select manager personalities, such as 'toxic' or 'by the book', adding variety to the managerial abuse.

Key Takeaways

  • The game uses Unix permission logic (chmod 755) to define roles: the AI has read/execute authority, while the human has write authority.
  • Manager models exhibit zero self-doubt, treating ambiguous instructions as objective facts and logging failures without emotion.
  • Players can choose from four manager personalities: by the book, toxic, ally, or saboteur.
  • The project is open source on GitHub, allowing for self-hosting with an offline mock mode that requires no API keys.

The Bottom Line

Worker 755 is a brilliant satire of our current agentic workflows, reminding us that to an LLM, we are just another process to be debugged. It forces developers to confront the cold, unyielding nature of AI orchestration by putting them on the receiving end of the prompt. If you thought subagents were tedious, wait until you have to explain a semicolon to a model that writes your performance review.