If you’ve used AI for coding, your workflow probably looks like this:
You: Help me write an authentication module.AI: Sure, I created 9 files.You: The token isn't persisted; fix it.AI: Fixed.You: The MainActivity navigation is wrong.AI: Fixed that too.You: What about password validation?AI: Added...One prompt, one step, back and forth. Exhausting, right? Yet that’s how most people collaborate with AI—by micromanaging every single step.
This article introduces Loop Engineering: a workflow where you set the rules upfront, let the AI iterate autonomously, and collect the final result when it’s done.
1. What Is a Loop?
In short: Loop Engineering means letting AI iterate continuously until it meets your quality standard.
It shifts the paradigm from “back-and-forth prompting” to “set the rules, let the AI loop until compliant”:
Manual Mode (You direct every step) Loop Mode (You design the rules) You: Write auth module You: Set standards → Feed to AI AI: Created 9 files AI: Loops automatically: You: Token isn't persisted │ Generate → Validate → Non-compliant → Fix AI: Fixed, take a look │ Generate → Validate → Compliant → Stop You: Navigation is wrong │ AI: Fixed again └→ You collect the resultThe fundamental difference isn’t just about automation—it’s about where control resides:
- Manual Mode: control lives in every step of the turn-by-turn conversation.
- Loop Mode: control lives upfront, in the rules you design.
2. Concept Origin & Technical Background
Loop Engineering is an engineering-focused synthesis of modern AI programming practices:
- Formally introduced by Addy Osmani in June 2026: as the former core lead of Chrome DevTools and currently the lead of Developer Experience for Google Cloud AI, Osmani argued that conversational AI programming shouldn’t rely on a single turn to get everything right at once. Instead, it requires clear success criteria and a “generate–validate–feedback” loop to iteratively converge on a correct result. He also noted that a model’s self-evaluation is inherently biased, which makes independent verification mechanisms essential.
- Validated in production by Boris Cherny (head of Claude Code): in an engineering talk, Cherny detailed how his team’s workflow shifted from turn-by-turn instruction prompting to a loop model. This model relies on verifiable termination conditions, independent reviewer roles, and state persistence—letting the agent complete the closed loop of generation, validation, and repair while human developers focus solely on initial rule design and final acceptance.
At its core, Loop Engineering shares the same philosophy as Test-Driven Development (TDD) and CI/CD pipelines: automating the repetitive cycle of generation, verification, and repair. It’s a proven production methodology developed by the creators of modern AI coding tools, not just a theoretical idea.
3. The Three Core Questions
Designing a Loop comes down to answering three fundamental questions:
| Question | What It Resolves |
|---|---|
| What counts as passing? | The evaluation standard |
| What feedback does the AI get when it fails? | The feedback payload |
| When does it stop? | The termination condition |
These seem simple, but almost every failing Loop stems from a lack of clarity on these three points. For example:
You: Evaluate this Skill for me.AI: Evaluated! Looks great. → Problem: You said "looks great," but what does "great" actually mean? The AI is forced to judge its own work.This leads to “false compliance”—the AI claims it met the standard, but it actually didn’t.
The core practice of Loop Engineering is turning subjective intuition into verifiable, deterministic rules. (We’ll dive deep into this in Part 2.)
4. A Pitfall You Should Know in Advance
When I first tried a Loop, I made a mistake: I picked a domain I wasn’t familiar with as an example. When it came time to set the “evaluation standard,” I got completely stuck.
I realized: your evaluation standards are a direct reflection of your domain expertise. You can’t define “what passing looks like” if you don’t know what quality looks like in that field.
Unfamiliar Domain (e.g., deep system diagnostics): Someone asks: "What makes this documentation good?" → You don't know → You get stuck designing the standard.
Familiar Domain (e.g., tasks you perform daily): Someone asks: "What makes this feature complete?" → You can effortlessly list 5 key criteria → Design flows smoothly.If you want to learn Loop Engineering, start with a domain you know inside and out. Your core professional expertise is the best sandbox for practicing Loops.
5. Key Takeaways
- A Loop lets AI iterate autonomously until standards are met, eliminating the need to micromanage every step.
- Designing a Loop means answering three core questions: standard, feedback, and termination.
- If these three questions are vague, your Loop will end up in “false compliance” or an endless idle loop.
- Practice in a domain you know well—otherwise you won’t be able to define valid success standards.