November 23, 2026AgentsTool

Worker 755 Makes You the Subagent, and the AI Writes Your Review

Everyone talks about orchestrators and subagents from the orchestrator's chair. Worker 755 sits you in the other one. It is a browser game where a real LLM, Claude, GPT, Gemini, DeepSeek, GLM or Kimi, runs a task queue with read-only access to a machine and a spawn subagent tool. The subagent is you. You get 90 seconds per task. It posted as a Show HN on Sunday and reached 242 points. https://github.com/gulivan/worker755 and https://worker755.com

What comes back to the manager is exactly what a real orchestrator sees: your report and the command transcript, or, if you froze, Subagent did nothing. Write a new instruction. Then it checks your work, files incident reports when you break things, and writes your performance review. The stakes are written into the fiction. A user named Dan is watching, and Dan's family photos and NDA are one rm away. Harder modes add a legacy production box with a live database, no command hints, and skipping failed tasks is off by default.

It is funny, but the joke lands because it is accurate. Sitting on the subagent side for ten minutes teaches more about orchestration than most framework docs. The instruction you receive is vague in exactly the way your own prompts are vague. The manager trusts your summary less than your transcript, which is the right instinct. And a confident one-line report after a half-done job reads very differently when you are the one who wrote it.

You bring your own OpenRouter or Vercel key, and the repo says a full shift on Claude Haiku costs a few cents at most. MIT license, about 105 stars, a weekend-sized project. Not infrastructure, just the best hands-on lesson available this week in why verification sits with the orchestrator and not with the worker.
← Previous
fast-anything Lets the Agent Write Rust, Then Makes Production Decide
Next β†’
app-store-compliance Gives Your Coding Agent an App Review Rehearsal
← Back to all articles

Comments

Loading...
>_