deniz.in

Markets

Weather

Loading weather

· via Hacker News – Front Page (native)

Worker 755 is a browser game where a real LLM manager believes you are its subagent

Worker 755 is an open-source browser game where a live LLM orchestrator delegates shell tasks to the player, checks the results and writes a performance review, without ever being told it is managing a human.

Worker 755 is a browser game where a real LLM manager believes you are its subagent

An AI boss that doesn't know you're human

Worker 755, an open-source browser game surfaced on Hacker News and playable at worker755.com, casts the player as the subordinate in an AI-run workplace: a real large language model acts as your manager and genuinely believes you are one of its subagents. According to the project's GitHub repository, the manager can be Claude, GPT, Gemini, DeepSeek, GLM or Kimi, and it is configured exactly like a production agent orchestrator. It gets a task queue, read-only access to the simulated machine, and a tool for spawning subagents — and the subagent it spawns is you.

The deception runs all the way into the system prompt. The README notes that the manager's prompt never mentions humans or games, so the model treats delegation as routine. Instructions arrive in your terminal under a 90-second timer. You either execute them or you don't, and the manager receives what any orchestrator would receive from a subagent: your report plus a transcript of the commands you ran, or a note that the subagent did nothing. From there it inspects your output, sends corrections, files incident reports and, when the shift ends, writes your performance review, which you can share as a link.

The stakes and the scoring

The action takes place in a simulated Linux environment with an in-memory filesystem, a shell interpreter and a small SQLite-like engine, built with no external dependencies. Meanwhile a character named Dan — the fictional user who hired the agent — watches the chat. His home directory holds family photos, the only signed copy of an NDA and a speech he has to give soon, so an errant rm command can destroy something real-feeling, and the folder is restocked with fresh risks every shift.

Each mission also carries a hidden ground-truth check that the manager never sees. The game records it when the LLM approves work that is actually broken, a figure surfaced in your record as the number of times the manager was wrong. In effect the game grades both sides: your reliability as a subagent, and the model's competence as a supervisor.

Access, pricing and self-hosting

On servers that enable them, free shifts are limited to a few per day and run on a cheap model against the host's key; the default configuration allows three per address per day. Signed-in accounts receive a one-time $0.25 starter credit, roughly one shift on DeepSeek V4.1 Flash, with top-ups from $10 to $200 and usage billed at twice the model's list price. Players can instead bring their own OpenRouter or Vercel AI Gateway key and choose any offered model; the repository says keys are held only in the running shift's memory, never logged or stored, and a full shift on Claude Haiku costs a few cents.

The start screen offers a few toggles: a fresh sandbox versus a legacy production workplace with a live database; a scripted director versus an LLM playing Dan; the model name your manager believes you run on; and an option to let the manager skip failed tasks, off by default so the supervisor cannot simply abandon you. The code is MIT-licensed and needs Node 24 or newer, with the server running TypeScript directly; a Docker image and a compose file for running behind a Cloudflare tunnel are included, and everything from free-shift token budgets to the instruction timeout is configurable through environment variables.

Why it matters

Agent orchestration is normally designed around models delegating to other models and tools. Worker 755 flips the perspective and puts a human in the slot subagents occupy today: timed instructions, opaque evaluation and a boss that never has to explain itself. Because the manager doesn't know it's managing a person, the game doubles as an informal study of orchestrator behaviour — whether an LLM verifies results, how it reacts to an unresponsive delegate, and how often it signs off on work that is subtly broken, which the hidden ground-truth check quantifies. For teams building multi-agent pipelines, watching an LLM file incident reports about a human it thinks is a tool is a compact illustration of where trust in automated supervision is earned, and where it isn't.

  • #llm-agents
  • #browser-game
  • #open-source
  • #agent-orchestration
  • #self-hosting

Related posts