TL;DR

The single biggest thing stopping engineers from adopting AI isn’t security, hallucinations, or code quality. It’s ego — the refusal to accept that an agent can do in thirty minutes what took them three days. Meanwhile the marketing coordinator, the ops person, and the analyst down the hall already reworked their whole week around these tools. The gate isn’t the model’s ceiling. It’s your willingness to be out-produced by it and keep your job description anyway.

Originally published at portfolio.hagzag.com.

The non-engineers already lapped you

Last month I watched a marketing coordinator — no code background, no CS degree — stand up a working lead-qualification pipeline in an afternoon. n8n, a couple of AI nodes, a Google Sheet. It wasn’t elegant. It ran. It saved her team a recurring day of manual triage every single week.

Two desks over, a senior engineer was still explaining, patiently, why he prefers to write his own boilerplate because “the AI doesn’t really understand the domain.”

That contrast is the whole post. The people with the least to prove are the ones moving fastest. The people with the most technical skill are the ones dragging their feet — and they’re telling themselves it’s about rigor.

In Part 1 I said the loop was the unlock. This is what blocks it.

The last post in this series — From Prompt to Loop: The Evolution of Agentic Development — argued that the real shift isn’t better prompts, it’s handing the model a loop and a harness and letting it iterate toward a measurable outcome. Prompt engineering, then context engineering, then loop engineering.

Every technical objection to that was answered a model generation ago. The blocker that’s left is human. You can give someone a loop that closes three days of work in thirty minutes and they will still find a reason not to press start — because pressing start means admitting the loop is better at the thing they built their identity on.

The thirty-minutes-versus-three-days problem is real, and that’s the problem

Here’s the uncomfortable part for a lot of my peers. I’m not making a motivational claim. In my own platform work I’ve watched an agent, running in a proper harness with tests and a definition of done, produce in one focused session what a competent engineer would bill roughly three days for. Scaffolding, IaC, the wiring, the first-pass tests, the docs.

Not always. Not on the genuinely hard, novel, judgment-heavy stuff — that’s still where humans earn their keep, and it’s a real category, not a face-saving footnote. But on the majority of day-to-day engineering work — the known-shape, been-solved-before work that fills most sprints — the gap is not close. And “three days into thirty minutes” is exactly the number that the ego cannot metabolize. A 10% speedup is a tool. A 20x collapse is an identity threat. So people reject the 20x and call the rejection “healthy skepticism.”

What the resistance actually sounds like

The tell is that the objections keep moving. Ship a model that writes clean code and the objection becomes “but it doesn’t understand our architecture.” Give it the architecture in context and it becomes “but I can’t trust it in prod.” Put it behind review and tests and it becomes “but reviewing its output takes as long as writing it myself” — which, notably, stops being true the third time you do it.

None of these are lies. Each one was true at some point. But watch the pattern: the goalpost moves exactly as fast as the capability does, and it always lands just past wherever the tool currently is. That’s not a risk assessment. That’s a person defending a self-image.

I did some of this myself. Three years in, the thing I had to get over wasn’t a technical doubt. It was the quiet feeling that if the machine could do the thing I was proud of being good at, then what exactly was I for.

What I’d tell the skeptic I used to be

The engineers who win the next few years are not the ones who out-type the model. That race is over. They’re the ones who move up a layer: deciding what to build, defining what “done” and “correct” mean, designing the loop, owning the outcome when it ships. The model does the typing. You do the judgment — and judgment is a bigger job than typing ever was, which is the part the ego actually wants to hear but won’t let itself.

The marketing coordinator didn’t beat the engineer because she’s smarter. She beat him because she had no professional identity riding on doing the work by hand, so she just… used the tool. That’s the entire advantage. It’s available to anyone willing to set the ego down for one afternoon.

What this costs if you don’t

The cost isn’t abstract. It’s compounding. Every sprint the skeptic spends hand-crafting known-shape work is a sprint the adopter spends on the judgment-heavy work that’s still scarce and still valuable. One of them is building the skill the market will pay for in 2027. The other is getting very good at a thing that’s being commoditized in real time.

Put the ego aside. Open the loop. Press start. You can pick the ego back up later — you’ll just be holding it somewhere higher up the stack.