July 14, 2026
Loops! Drama! Trench coat!
The Agentic Loop: Three loops in a trench coat
One simple AI idea turned into a comment-section fight about circles, diagrams, and endless loops
TLDR: The article says today’s AI helpers are really built from three repeating steps, not one simple magic process. Commenters instantly turned that into a fight over whether the diagram is backwards and whether there are actually endless loops, showing how messy — and important — these systems are to understand.
A fresh post about how so-called “autonomous” artificial intelligence really works was supposed to simplify things: instead of one magic process, the author says it’s three separate cycles stacked together — the machine thinking, the machine using outside tools, and the human stepping in. In plain English, the point is that these systems aren’t one smooth robot brain. They’re a bundle of repeated steps wearing a disguise. Cute metaphor, clear mission — and then the comments immediately did what comments do best: started a tiny nerd civil war.
The fastest pushback came from readers squinting at the diagram and asking, wait, are the circles even in the right order? One commenter bluntly argued the smallest, fastest cycle should be the machine’s word-by-word thinking, with the tool-using cycle outside that, and the human cycle outside both. Then swyx kicked the door open with the classic internet escalation move: actually, there are more than three. That invited the funniest reply of the bunch, with tptacek basically saying if you want to keep blogging, you can always invent more loops because, at some level, all software is just one giant loop anyway.
So the real mood here wasn’t outrage — it was playful tech pedantry at full volume. The jokes landed, the trench-coat image stuck, and the crowd turned a practical explainer into a surprisingly dramatic debate over whether AI is three loops, many loops, or just loops all the way down.
Key Points
- •The article argues that AI agent behavior is better understood as multiple nested loops rather than a single loop.
- •Ross defines the inference loop as responsible for chat completion API calls, passing tool requests onward, and persisting chat history.
- •The article states that most LLM APIs are stateless, so developers must resend the full conversation history with each request.
- •Ross says tools are what give LLM-based systems functional agent behavior.
- •The article notes that models may request multiple tools in one turn, which is why a separate tool loop is needed.