How to run a Gauntlet Loop: pair each AI builder agent with a blind critic

Matt Shumer has named the method behind Claude of Duty, the browser shooter he says Claude Opus 5 built from one prompt. In a July 27 how-to on his site, Something Big Is Happening, he calls it the Gauntlet Loop: every piece of the work gets a builder agent plus a separate critic that judges it against a real reference.

On July 25, Shumer (@mattshumer_) posted a video on X of a Call of Duty-style first-person shooter. He wrote that Claude Code ran unattended for hours, used a large swarm of subagents and wrote roughly 55,000 lines of code, generating all art and audio in code. When viewers called it fake, he released the prompt and the Three.js source on GitHub (mshumer/Claude-of-Duty).

How it works

• Use an agent harness such as Claude Code or Codex that can render and inspect its own output, not a chat window.
• Give the lead agent the goal and a concrete bar it can open. For the game, that was real Call of Duty screenshots.
• Let the lead agent break the goal into the smallest units a critic can score on their own, such as the gun, hands, trees, lighting, movement or enemy behavior.
• Give each unit a builder and a fresh-context critic that never sees the builder's reasoning. The critic compares the work and the reference blind and picks the better one; if the reference wins, it names the biggest gap and sends the work back.
• Don't fix a number of rounds. His game was still getting better when he halted it.

Why it matters

The key is separation. Shumer argues a builder knows why it made each call, so it's good at defending its work and shouldn't grade it. Our earlier story on a 61 KB WebGPU ocean covered a related approach: an agent that screenshotted its own work, and SKatalystAI's variation where a fresh critic agent reviewed a 52 KB neon city three times.

Others have packaged it. The duolahypercho/gauntlet-loop repo turns Shumer's prompt into a slash command for Claude Code and Codex; its README warns the loop never ends on its own. Yash-1511/gauntlet-loop is stricter: it keeps evidence in a .gauntlet/ folder and ends a run as PASS, BLOCKED or PLATEAU. Favur Dev (@favurdev) wrote on X on October 3 that its harness now runs critics in code, counting a score only if a screenshot or test output backs it.

Limits and caveats

AIGamer hasn't run a Gauntlet Loop, and the results above are the authors' own. In an August 12 WotAI blog post, Alex Kim reported testing the method on four projects and found the biggest predictor of a useful run was whether the critic had something real to inspect. "Adjectives get talked around," he wrote. One of his rounds was discarded because Chrome dropped part of a screenshot. He also warns that runs take hours and a real budget.

How to try it

Put real reference images on disk, have the lead agent keep a live progress page of screenshots, and pick a stopping rule in advance. Shumer suggests stopping when you like the result, when gains get small or when the compute budget runs out. His page also includes a meta-prompt that writes a Gauntlet Loop prompt for your own goal.