01What Are Claude Agent Loops and How Do They Work?

Claude agent loops are systems where an AI agent acts, a separate judge checks its work against a goal, and the loop repeats until the work passes quality control—all without you prompting at every step. In this episode, Julian Goldie explains five different loops he built inside his agentic operating system (AgentOS).

Instead of typing, reading, and correcting every turn like in a chat, you write the loop once. You define what 'done' looks like. Then the agent acts, observes, judges, and repeats on its own.

It runs while you sleep, and you wake up to finished work.

02The Five Loops: Building, Planning, Shipping, Watching, Writing

Julian built five loops that all run on one shared memory. Each one has a different purpose:

  • Loop Engineering: Build anything to a spec, like a website, with a builder and a judge.
  • Kanban Board: Uses three roles—planner, builder, reviewer—to move tasks from triage to done automatically.
  • Pipeline: Turns an idea from inbox to shipped, with one approval point.
  • Oracle: Watches the news every morning and hands you the six biggest AI stories to post about.
  • SEO Pipeline: Generates keywords, writes SEO articles, and publishes them with one click.

The memory system connects all loops, making them smarter over time. It knows your projects, team, vision, voice, brand, and companies—all stored in a 'memory galaxy.'

03How to Set Up Claude Agent Loops (If You Want to Build Your Own)

Julian explains that you can build loops yourself. He shows the steps for loop engineering:

  1. Define what 'done' looks like for your project.
  2. Give the agent a starting point if you have one.
  3. Select the builder API (for example, a free one to save tokens).
  4. Set how many rounds the loop should run.
  5. Choose the judge API (for example, a powerful model like GLAM 5.2).
  6. Let it loop round until the work passes quality control.

You don't need to touch anything. The loop saves outputs to your workspace, so you can check them later. If you don't want to wire up five systems, Julian offers the full AgentOS inside the AI Profit Board, which you can give as a zip file to Claude or Hermes to set up.

04How to Use the Loops: Examples from the Episode

Julian walks through each loop with real examples:

Loop Engineering: You type 'build a website' into the system. The builder creates it, the judge checks it, and it loops until done. You can see the outputs saved in your workspace.

Kanban Board: In a content system, you drop a task like 'create a video' into the board. A video builder, a director, and a content judge work together. The task moves automatically through triage, to do, ready, running, blocked, and done.

Pipeline: You capture an idea, an agent classifies and plans it, then you approve once. A project manager and sub-agents build it out. The whole thing lives in a vault so nothing gets lost.

You can delete ideas you don't like.

Oracle: Every morning it searches for the latest AI news, ranks it by relevance, and gives you six stories with original sources, your unique angle, and a ready hook. One click turns a story into ranked SEO articles.

SEO Pipeline: It generates keywords, you click 'use topic,' then it generates the article and publishes it—all in one click.

05Results and Comparisons: Old Way vs. Loop Way

Julian compares the old prompting method to the new loop system. The old way makes you the engine: you type, wait, read, correct, and repeat. Nothing moves unless you're at the keyboard, and your output is capped by your attention.

The new way makes you the architect: you write the loop once, define done, and the loop runs itself. It acts, observes, judges, and repeats on its own. You wake up to the work done.

For results, Julian mentions that over 196 pages of testimonials and wins exist from people using the AgentOS inside AI Profit. He also shows the Oracle was last consulted four hours ago, and he's published loads of content automatically. The memory system updates every few hours, making the loops compound.

06Tips, Limits, and Mistakes to Avoid

Julian shares several practical tips:

  • Use a powerful model like GLAM 5.2 for the judge, and a free API for the builder. This saves tokens and costs less.
  • You can add a human in the loop for a quick check before things go out.
  • You don't need to be technical. You can give the zip file to Claude or Hermes and it will set up for you.
  • Make sure you define done clearly, because the loop only knows what you tell it.

One limit he highlights: with the old way, you were the loop, and your output was capped by how long you could stare at a screen. With loops, you still need to approve plans in the pipeline, but you can delete ideas if you don't like them.

07Who Should Use Claude Agent Loops and Why It Matters

Claude agent loops are for anyone who wants to get more done without sitting at the keyboard. Julian says he's non-technical, so if he can build these, you can too. The loops run in the background, so you can focus on being the architect rather than the engine.

Why it matters: you can build, plan, ship, watch, and write—all on autopilot. The shared memory makes every loop smarter over time. This saves a lot of time and lets you handle many projects at once.

If you want to use them, the AgentOS is available inside the AI Profit Board, with coaching calls, community support, and daily guides.