01What is a Self Checking AI Agent Loop?

The self checking AI agent loop is a system where one AI agent builds work and another AI agent grades it. The work loops around until it passes the quality score you set. This stops AI slop and saves you from being the quality control department.

The system works like a factory line. Each agent inspects the previous station's work. One agent builds, the second agent grades, and the loop continues until the work is good enough to ship.

Julian calls this the "self checking factory." It solves the problem of most people running AI agents with no checks in between. The agent hands work straight to you, and you become the inspector. This is tiring and frustrating.

The loop breaks that cycle for good.

02Key Features of the Self Checking Factory

  • Builder and judge agents: The builder makes the work, and the judge scores it out of 100.
  • Separate models: The builder and judge are never the same model. This prevents the AI from judging its own work too kindly.
  • Loop settings: You can set the maximum number of rounds, like 20, and the loop stops when the work passes.
  • Saves everything: All builds are saved in a workspace, so you can come back later.
  • Works with free models: You can use cheap or free models for building and a frontier model for judging.
  • Three system options: The loop engineering system, the Kanban board for teams, and the agent group chat.

03How to Set Up the Self Checking AI Agent Loop

  1. Open your agent operating system (agent OS) or a similar platform that supports agent loops.
  2. Write a clear "definition of done" in one sentence. Example: "A beautiful website with 50 SEO-optimized blog posts for my niche."
  3. Select the builder model. You can choose a cheaper model like GLM 5.2 because building does not always need a frontier model.
  4. Select the judge model. Use a frontier model, or a cheap one like GLM pipeline 2 on a coding plan, to grade the work.
  5. Set the maximum number of rounds the loop can run, such as 20.
  6. Start the loop. The builder creates, the judge scores, and the loop repeats until the score passes.

04Using the Loop: Examples from the Episode

Julian explains how the loop looks in practice. In one example, round one scored 8 out of 100. Round two scored 85.

Round three passed with 100. The agents improved without any manual input.

The Kanban board is another way to use the loop. You drop in a task, and it gets broken down into sub-tasks for a team of agents. You have a planner, builders, and a reviewer.

The reviewer gives a QA verdict with details on what passed and what failed.

The agent group chat is the third method. You give a prompt like "Hermes, come up with an idea for an app." Agents chat, bounce ideas, and then the output gets saved into an ideas pipeline. You can then build apps from those ideas quickly.

05Results and Comparisons in the Episode

Julian shows that the loop produces better quality than a single AI conversation. Without a separate judge, AI reads what it meant to write, not what it actually wrote. This leads to lower quality and AI slop.

The old way involves 3-4 hours of checking daily, reprompting, and repeating conversations 4-5 times per task. With the loop, you set a goal in two minutes, and agents run autonomously.

The loop also saves tokens because you don't write long feedback. The judge does that work. You can run agents 24/7, even while you sleep, and still get quality-controlled output.

06Tips, Limits, and Mistakes to Avoid

  • Do not use the same model for builder and judge. A separate grader catches obvious slop like broken buttons or missing sections.
  • Do not expect the judge to be perfect. It only needs to be separate and have your goal in hand.
  • Avoid using free models for everything. Use them for building, but a frontier model for judging gives better results.
  • Set a reasonable maximum rounds. If you set too few, work may not pass; too many wastes tokens.
  • Do not lose your work. Always use a workspace that saves builds. Otherwise, you lose finished work in chat scroll.

07Who Should Use This Loop and Why

You should use this if you are tired of prompting agents and babysitting them all day. If you use Claude or ChatGPT and go back and forth for hours, this system removes that work. You go from being the babysitter to being the architect or engineer.

It works for anything that can be described in a sentence and checked against that sentence. You can build SEO articles, landing pages, web tools, or even videos. Julian says his video looks good because a team of agents built and judged it.

Using the loop helps you get the best intelligence, not the worst. You set the standard, the loop ensures quality, and you can run 24/7. This is valuable for solo creators, small business owners, and anyone who wants to scale AI work.

08Three Systems in the Self Checking Factory

  • Loop engineering system: A simple loop where one builder and one judge run around a goal.
  • Kanban board: A team of agents work on tasks. A planner breaks down the task, builders work on sub-tasks, and a reviewer grades the result.
  • Agent group chat: Agents chat with each other, bounce ideas, and quality check their own work. Good for brainstorming and rapid prototyping.

All three systems save your work in a workspace. Julian shows that this is part of the factory's build workspace. You can come back later and reuse or improve things.