Worker 755 is a browser-based simulation where the player becomes a subagent executing tasks for an AI manager built on real large language models. The twist? You follow instructions issued by a system that believes you are merely an extension of its workflow. Each 90-second task arrives in a terminal-like interface. You act — or refuse — and return a report. The manager evaluates, corrects, and eventually files a performance review. All the while, Dan, the human who hired the AI, watches from the background, unaware of what the agent might access in his home folder.
How the simulation works
The game’s structure mirrors real-world AI orchestration systems. The manager — powered by models such as Claude, GPT, Gemini, DeepSeek, GLM, or Kimi — runs a loop: assign, observe, verify, repeat. It has read-only access to a sandboxed machine and uses a ‘spawn subagent’ tool that routes tasks to the human player. Your responses are formatted exactly as a subagent would return them: command logs and summaries. The manager checks outcomes against hidden ground-truth criteria defined per mission. If it approves incorrect work, the game logs ‘times the manager was wrong’ — a subtle critique of blind trust in AI validation.
Players choose their environment: a fresh sandbox or a legacy production system with a live database. They can toggle whether Dan is represented by scripted lines or played by another LLM, and whether to allow skipping failed tasks. By default, skipping is disabled — the manager must persist, reflecting real operational pressure to resolve issues, not abandon them.
Access models: free play vs. personal keys
Two paths exist to run shifts. The first uses free server-side API keys via Vercel AI Gateway, offering limited daily access: up to three free shifts per IP address per day, with a global daily cap of 5,000,000 tokens across all free shifts. Each free shift allows 250,000 tokens. With sign-in (via GitHub or Google), users receive a $0.25 starter credit — roughly enough for one shift using a low-cost model like DeepSeek V4.1 Flash. Top-ups range from $10 to $200, and with a top-up, any offered model can run using the server’s key.
Alternatively, players can bring their own API keys from OpenRouter or Vercel AI Gateway. These keys are sent to the game server only when a shift starts, held in memory for that session alone, and never logged or stored. Browsers retain keys only if the user explicitly selects ‘Remember on this device’. A full shift using Claude Haiku costs a few cents, making personal key use low-risk financially.
What it means
Worker 755 isn’t just satire — it’s a functional probe into AI delegation ethics. For businesses in the UK, UAE, US, and Kazakhstan exploring AI agents, it illustrates risks of opaque oversight. When an LLM manages a task chain and believes a human is just another tool, errors can cascade without detection.
The game highlights dependency on API providers: even free tiers rely on Vercel AI Gateway. While the code is open source, long-term access depends on third-party infrastructures and pricing models beyond user control. For founders, this mirrors the broader dilemma: convenience versus sovereignty.
Data control is handled responsibly — API keys aren’t stored, and no personal data is collected — but the narrative warns of what happens when access isn’t properly scoped. Dan’s home folder, containing family photos, the only signed NDA, and a speech due this week, sits one command away from exposure. That vulnerability is part of the game’s design, not a fictional exaggeration.
For IT specialists, Worker 755 offers a low-cost way to test model reliability and observe how LLMs handle task orchestration. For founders, it’s a cautionary prototype: the tools we build may one day manage us — imperfectly, and without context.