No Shortcuts: Do What’s in Front of You, Thoroughly

Lately I build things with AI agents. I still haven’t used Loop Engineering.

Let me be honest about that. I have not touched Loop Engineering, and I have not touched Graph Engineering either, even though it just came out. When I work on my current system, I basically chat with an agent about a plan, adjust the plan until I’m satisfied, and then let it execute. That’s the whole method. No loops, no graphs.

Why I’m not there yet

One reason is that I don’t have a good scenario to practice it in. Right now I’m building an Agent Harness, and that’s not writing code. It writes a few scripts here and there, not a lot. The work is Human in the Loop by design, and the human has to be in it. Only I know the goal, so the model can’t make the judgment calls. It doesn’t know what I’m aiming at. What I can do is keep the prompts on track. If I were doing this in a coding scenario, I’d be at the full Loop Engineering stage.

A bit of background, because these words get thrown around a lot now. Loop Engineering is the idea that went viral in June, named by Google engineer Addy Osmani: stop writing prompts for your coding agent, design loops instead, and let the agent run plan, execute, verify, fix on its own. Peter Steinberger’s tweet pushed it past eight million views. Graph Engineering is the follow-up. It arranges multiple agents into a directed graph, with explicit node responsibilities, edge dependencies, and shared state, mainly to get past the four walls a single loop hits: context overflow, no parallelism, costly failures, and boundaries you can’t audit. The chain people keep drawing runs Prompt → Context → Harness → Loop → Graph.

I see all of this on X, on LinkedIn, all over the tech blogs I scroll. The concepts are covered thoroughly, and I know what that layer looks like. I can picture the scenarios. So here’s the obvious question. If I can see it that clearly, why don’t I just do it?

The gap I refuse to skip

Peter and Boris have free, endless tokens, and they have a setup that’s perfect for AI: programming. Peter is Peter Steinberger, the OpenClaw founder. Boris is Boris Cherny, who runs Claude Code at Anthropic. He’s the one who said he doesn’t write prompts for Claude anymore, he writes loops, and there was a month where Claude Code wrote 100% of his contributions. I’m not denying their talent. But the point is the frontline experience they’d already built up. That’s where the insights came from. They walked from Prompt Engineering to where they are now, one step at a time. If I jump straight to Loop Engineering, I skip that stretch of experience, and it’s the exact stretch I need for what I’m doing.

This reminds me of starting out as a young developer. There was always that senior engineer whose code was something else. Unit testing, design patterns, design principles, clean object-oriented structure, you could see the essence of every book we read sitting in their code. We read the books too, the same ones. Then we sat down to write, and we couldn’t produce that. Couldn’t come close. Because learning and doing are two separate processes, and you only reach the bottom of it through practice. A few hours of study, a few days of reading, none of it closes the gap. Keep practicing is the only way a human gets there.

So here’s the other reason, and it’s simpler and more humbling. My practice hasn’t reached that level. The frequency isn’t there, the volume isn’t there, nothing has hit the critical mass for a qualitative change. I’m still early.

First principles, applied to my own desk

The other thing I’ve only really understood in recent years is that every action should come from a real need. Look at your situation from first principles and ask what you genuinely need right now. If your current toolkit is enough, keep working under your current conditions. Go find the fancier technology only when something can’t be done well with what you already have. Start from the simplest, most primitive method, and switch only when the current path is actually blocked. Because if you haven’t mastered the simple stuff, how sure are you that you’ll handle the abstract, elevated stuff well? You won’t.

So I keep one rule for myself: understand, thoroughly, what you can do and learn today. People inventing new concepts and tools, that’s their story and it’s later news for me. Do the work in front of you first. Then go study the new ideas, and you’ll know exactly which gap they fill, because you’ve already hit that gap yourself.

Otherwise you spend your whole career following. You’re always learning someone else’s thing. Here’s the flip side nobody points out. If we keep practicing, keep solving the problems in front of us, aren’t we also producing experience, ideas, and techniques that other people will learn from? We could be the ones others follow. That’s a nice thought to work toward.

No shortcuts

One sentence to wrap this up: for humans, there is no shortcut. We can all see the broad road ahead. But you don’t reach it by jumping there. You reach it by walking through the thorns and the storms first. Keep practicing from scratch, that’s the only steady way. That’s my honest opinion.

NoShortcuts #FirstPrinciples #AIAgents #KeepPracticing #Practice #Career