A new open-source browser game called Worker755 reverses a familiar agent setup: an AI model acts as the manager, while the human player occupies the role the model believes belongs to a software subagent. The result is a timed simulation built around delegated terminal work, machine access and the sometimes unreliable judgments of an automated supervisor.

According to the project repository, the manager is configured like an orchestration agent with a task queue, read-only access to a machine and a tool described as spawning a subagent. When that tool is used, its instruction appears in the player’s terminal with a 90-second countdown. The player may carry out the task or refuse. The manager then receives the player’s report and command transcript, or a message saying that the subagent did nothing.

The interaction continues beyond a single assignment. The model can inspect the reported result, issue corrections, create incident reports and eventually write a performance review. A fictional user named Dan watches the conversation while the game places personal-looking files in a home directory, creating stakes around careless commands. The repository says each mission contains a hidden check, allowing the game to record cases where the manager approves work that is actually broken.

Worker755 supports several model families through its manager role, including Claude, GPT, Gemini, DeepSeek, GLM and Kimi, but the hosted or self-run configuration determines which services are available. The project does not conceal that operational requirement: its documentation says players normally bring their own model access unless a deployment operator provides free shifts through a configured gateway. It also includes an offline mode with a scripted manager and no network connection.

The software requires Node.js 24 or later for local use and runs its TypeScript server directly. The repository includes a development setup with a Vite front end and mock server, plus a container deployment configuration. Its suggested Cloudflare arrangement uses an unprivileged, read-only container, resource limits and no published ports. Shared performance reviews can be stored and distributed as links after a shift.

The game’s central joke depends on a real ambiguity in modern agent systems: from the manager’s perspective, a concise report and command log may look the same whether produced by another model or by a person following instructions. Worker755 turns that abstraction into the play mechanic while exposing failure modes such as vague delegation, misplaced confidence and approval without adequate verification. It is an entertainment project rather than an evaluation benchmark, and its hidden mission checks do not establish how any model performs outside the game. Even so, the design offers a compact demonstration of how tool-using agents assign work, observe limited evidence and decide whether a task is complete.