Issue #3 — Loop Engineering (Jason Zhou)

This is the first installment of a new series. Each issue profiles one builder's framework for working with AI agents — the named systems people actually run, not the theory. Coming up: IndyDevDan's hooks and skills, Nick Saraev's Claude Routines, Cole Medin's Archon harness, and Matt Van Horn's /last30days skill.

Remember when the skill was writing a good prompt? That was 2023. Then came parallel agents — firing off five agents at once and herding them. That's where most people stalled out, because herding five agents feels like babysitting.

Jason Zhou says the next step is loop engineering: stop babysitting prompts and start designing the loop that runs the agents.

Zhou runs AI Builder Club — a learning community with weekly live builder workshops — and his YouTube channel (~231K subscribers) walks through real agent builds. In June he released his video "wtf is Loop Engineer & how to setup for real" and an open-source loop-engineer template, and AI Builder Club published its "Loop Engineering Guide (2026)." The core idea is simple: an agent doing one task is a tool. A loop that runs agents, checks their output, feeds it back in, and runs again is a system. The skill isn't the prompt anymore. It's the loop design — the checkpoints, the feedback paths, the stop conditions.

He's not theorizing. In May he published "Claude Code for Startup Founders: Ship Your MVP Fast" — a 5-day idea-to-MVP workflow built on this thinking. His whole lane is translating agent work for people who aren't engineers: the idea→spec→MVP translator. If the framework only worked for researchers, it wouldn't matter. It works for founders with a deadline.

Zhou lays it out as three stages:

The money move is in stage three. A loop that works keeps working overnight, unattended, on real tasks. That's the difference between a demo and an employee.

Zhou's been productive this year, and the adjacent work is worth knowing. This summer he took on "graph engineering" (a July write-up, then an Aug 2 workshop, "Graph Engineering Without a Framework") — how to structure agent workflows as graphs instead of chains, so work can branch and merge. And in a Sept 20 workshop he covered Jev, the pattern now called "Jev Engineering: LLM writes, Jev decides, code acts" — splitting the job into three roles so the language model does the creative work, a decision model (Jev, from TypeSafe AI) makes the calls, and code does the acting.

The thread connecting all three: separate the thinking from the deciding from the doing, then loop it. That's the framework in one sentence.

Each upcoming issue takes one framework from one builder and breaks it down the same way:

You don't need to become an AI engineer. But you do need to understand the shift Zhou is describing, because it changes what you buy.

When you hire someone to "build an AI thing" for your business, ask whether they built you a prompt or a loop. A prompt needs you to run it. A loop runs itself and checks its own work. One is a tool you babysit. The other is closer to an employee.

At CED, we're adopting the best of these primitives into our own build factory — the loops, the checkpoints, the stop conditions. Every framework in this series gets evaluated against one question: does it make our client builds more reliable? If yes, it goes in.

If you want to see what loops could run inside your own business — which tasks are loop-shaped and which still need a human — that's what our $97 AI-Readiness Audit maps out. Written breakdown, gaps first, fixes second.

Sources: Jason Zhou (founder, AI Builder Club; AI Jason YouTube, ~231K subscribers — "wtf is Loop Engineer & how to setup for real," June 18, 2026, and the loop-engineer-template repo; AI Builder Club, "Loop Engineering Guide (2026)", June 17, 2026; graph engineering LinkedIn post, July 2026, and workshop "Graph Engineering Without a Framework," Aug 2, 2026; workshop "Jev: Decisions Instead of Text," Sept 20, 2026, per the AI Builder Club workshop archive; "Claude Code for Startup Founders: Ship Your MVP Fast," May 2026). Series installments: IndyDevDan (Tactical Agentic Coding; claude-code-hooks-mastery repo; "My Super Simple Software Factory," Sept 2026; "agents plus code beats agents alone"); Nick Saraev ("Claude Routines Just Dropped, And It's Perfect," Apr 14, 2026 — email triage, transcript-to-proposal, HackerNews scraper; routines beat n8n); Cole Medin (Archon rewritten June 2026 as "the first open-source harness builder for AI coding"); Matt Van Horn (/last30days skill — MIT-licensed, research across Reddit/X/YouTube/TikTok/HN/Polymarket/GitHub; works in Claude Code, OpenClaw, Cursor, Gemini CLI; Phala Cloud prebuilt template).

From the Human-N-Loop shop

Original AI-art gear from our own store — three picks for this issue:

Human-N-Loop Metal Pin – Teal Bot — $24.99

Human-N-Loop Water Bottle – Robot Trust Me — $44.99

Human-N-Loop Dad Cap – Type Stacked — $29.99