What is a Gauntlet Loop? Building Games with Agent Critic Loops

What is a gauntlet loop?

A gauntlet loop is a method of developing games, apps, and other projects using AI agents in which a separate critic agent repeatedly critiques the builder agent’s work in comparison to a fixed reference. This produces high-quality results through many rounds of revisions, rather than settling on the first passable result.

Gauntlet loops were first introduced in July 2026 by Matt Shumer, who posted on X about building a Call of Duty-style shooter in Three.js with a single prompt. He coined the term gauntlet loop and wrote about it on his blog.

How does a gauntlet loop work?

A coding agent, such as Claude Code or Codex, is given a goal and an example of what the finished work should look like. This lead agent performs the orchestration, breaking the project into subsections that can each be built and critiqued by sub-agents. For each piece, the builder produces an attempt at it, then the critic compares it to the provided reference for complete work; if the reference is better, the critic gives the builder feedback and the loop continues.

There are a few key principles to achieving the best results with a gauntlet loop, which include:

Don’t describe the implementation details

Agents are good at deciding how to build something, and they’ll be able to make judgments based on their architecture for the project. Describing the details of what you want to build is important, as this is where your creative vision comes in, but try to avoid describing how to implement it.

Give the agent a specific, verifiable bar

Having a clear, high-quality reference will make it easier for your agent to improve its work. Screenshots are great, or specific websites, written pages, or whatever fits what you’re trying to build. It should be easy for the agent to compare its work to your reference and determine which one wins. The bar doesn’t actually have to be achievable, but the stronger of a reference the agent has for comparison, the better results you’ll get.

Let the agent orchestrate

As the agent builds the project, it will be able to understand what tasks are left, which specific parts need their own verification loops, and where the bottlenecks are. Besides the initial instruction to use critic loops, leave the agent to decide the specifics of where to distribute sub-agents and create sub-loops.

Only let the critic agent judge

This is perhaps the most important part of the loop. The builder agent knows what it created, why it made the decisions it did, and how the project is structured, so it is unfair to allow it to judge its own work. The critic should be modeling a consumer, who wants a high-quality product and doesn’t know or care about the details of how it was made.

Don’t limit the loop

The gauntlet loop should not have a deterministic “final round.” Rather, it should run until the critic agent is satisfied and believes the game cannot improve any further. Even if you hit a session limit while building, you can simply resume the loop after your limit resets. After the agent has provided a final result, you can continue to issue follow-up prompts for additional changes and improvement.

How to improve asset quality in gauntlet loops

One of the most common walls encountered in gauntlet loops is that the quality of procedural 3D assets plateaus, while the critic keeps reporting that the environment is weaker than the reference. Once this happens, it becomes difficult for the builder to keep improving, as procedural assets can only go so far.

Thrixel generates high-quality, game-ready 3D assets with named, editable parts, so the agent can refine and edit individual models piece-by-piece, instead of having to regenerate from scratch when something isn’t right. This fits the gauntlet loop workflow perfectly, as it enables the agent to perform the same work with 3D assets as it does with the rest of the project: keep comparing to the reference, and iterate until the critic is satisfied.

Running gauntlet loops with Thrixel’s Build World skill

Build World is Thrixel's open-source agent skill for the gauntlet loop workflow. With the skill installed, you can describe any game, environment, or other 3D world, and the agent builds it start-to-finish using Thrixel for high-quality and editable assets. Further, you can access all the generated assets in your Thrixel account, so you can make further edits and have the agent pull the updated assets into the game.

Get the skill repo

Common questions about gauntlet loops

Is a gauntlet loop the same as vibe coding?

They overlap. Vibe coding or vibe gaming is the general practice of building a game by describing what you want to an agent. A gauntlet loop is a specific technique for achieving high-quality results from vibe coding or gaming using the builder/critic agent checking paradigm.

Do I need a particular tool?

You need an agent that can run code, render the result, look at it, and spawn other agents. A plain chat window won't work, because the critic has to be able to inspect the builder’s work. To use Thrixel for assets, you can create a free account here.

How long should a loop run?

Longer than feels reasonable. With a high enough bar, there's almost always room for improvement, and just because a game is passable doesn’t mean the agent can’t keep improving it.