The gauntlet loop is an agentic prompting pattern in which an AI agent system iterates on a piece of work against an explicit quality bar, using separated builder and critic roles and blind comparisons, until the output beats a concrete reference. The technique was popularized in 2026 by developer Matt Shumer, who coined the name and used it to build "Claude of Duty", a browser game produced end-to-end by agents inside an agentic coding harness.
How it works
A gauntlet loop starts from two inputs: an ambitious goal and a concrete example of what "great" looks like (the reference). A lead agent then:
- Sets the bar. The reference is treated as the standard to beat, not as inspiration.
- Decomposes the artifact into the smallest parts that can be improved and judged independently.
- Assigns builders. Each important part goes to a builder agent that produces or revises it.
- Judges blind. A separate critic agent with fresh context compares the current output against the reference as a blind A/B test, without knowing which is which.
- Finds the largest gap. If the output loses, the critic names the biggest meaningful difference, which becomes the next build target.
- Loops until the output wins the comparison or a stopping condition is reached.
The separation between builders and critics matters: a fresh-context judge avoids the self-grading bias of a model evaluating its own work, and the blind A/B framing forces concrete comparisons instead of generic praise.
Requirements and uses
Gauntlet loops run inside agentic harnesses such as Claude Code or Codex, where agents can open files, run code, render results, inspect screenshots and spawn sub-agents. The pattern applies to any artifact that can be inspected and improved: code, websites, product design, marketing assets, writing and research. It belongs to the broader family of agentic self-improvement techniques alongside critic-verifier pipelines and judge panels in multi-agent systems.