AI Pulse · Under the Radar

Matt Shumer dropped a single prompt into Claude that built a working Three.js first-person shooter with no hand-coded assets. He called it “Claude of Duty.” Within three days the repo hit 1,547 stars and 253 forks, the fastest velocity for a new repository in this cycle. That pace tells you something: builders are already running their own versions.

One operator let Opus 5 work solo for 24 hours and came back to a complete game. Another spun up what they’re calling an “adversarial subagent” pattern, crediting Shumer’s original prompt structure. Shumer is now naming the underlying workflow the “Gauntlet Loop.” The idea is simple. You pair a specialist builder agent with a ruthless blind-critic subagent that only approves output if it beats a real-world benchmark. The critic does not see the builder’s process, only the result. If it is not better than what already exists, the loop rejects it and the builder tries again.

This is not vaporware. It is a working pattern that people are forking and testing right now. If you run any kind of content pipeline, proof-of-done review, or design iteration inside your company, the Gauntlet Loop is worth trialing before someone wraps it in a product and charges you monthly for it. The core idea, a builder that never ships until a blind critic says it clears the bar, fits naturally into the kind of multi-agent orchestration we build into the Omni Command Centre. You do not need to wait for a SaaS tool. You can test the pattern this week with your own agents and your own quality threshold.

The speed of the copycat wave matters more than the game itself. When a workflow spreads this fast, it means builders see immediate utility. That is the signal.