Part 1 — The Education Loop
Developing solutions for over 15 years, you often reach a point in your projects where you wish you would have done things differently. Then you do things differently in the next project, and down the line you feel the same way. This is the self-taught education loop of engineering real-life solutions.
Nobody frames it this way, but the lessons from the loop aren't technical. Sure, you learn to stop putting business logic in your controllers or to stop treating your database like a junk drawer. But what actually compounds is the instinct. You start feeling when something is wrong before you can say why. You walk into a codebase and something in the structure bothers you and three weeks later you find out you were right. That doesn't come from tutorials. It comes from having built the wrong thing three times and sat with the consequences long enough to internalize what went sideways.
Now, with AI, you can collapse those loops faster. You don't have to wait until the next project. You can spin off a complete parallel rebuild around your new thinking. Don't let the sunk cost fallacy drag you into the mire. Keep iterating, keep learning, keep going.
AI can speed up the building. It can't give you the instinct. That still has to be earned the hard way.
Part 2 — The Parallel Rebuild
Here's what changed. Before, the cost of being wrong was the rest of the project. You'd see the better path six months in and just live with it because tearing it out would set you back further than pushing through. Now you can fork reality. Take everything you learned mid-project and spin up a parallel version that does it right, or at least differently, while the original keeps running. The feedback cycle used to be project-to-project. Now it's week-to-week. Run the version that works alongside the version that's better until you're confident enough to cut over.
The sunk cost argument stops making sense when starting over costs a weekend instead of a quarter.
This is the thing to sit with. Not the speed. The permission. Permission to be wrong and act on it immediately instead of filing it away for the next project. Fifteen years in and I'm still finding things I'd do differently. The difference now is I can actually act on that in real time.
Part 3 — Agents Don't Sleep
I built Hermes to handle the work that doesn't need me. Sensor fleet analysis, forecasting, nightly batch processing. The stuff that used to eat my mornings before I could even think about building anything new. Now I wake up and the analysis is done. The issues are filed in GitHub. The boring infrastructure of running a real system just runs.
The gap between a demo and a product is all the unglamorous stuff nobody wants to talk about on Twitter. The monitoring. The 2 AM pipeline failure. AI agents won't close that gap for you, but they can hold the line while you're offline. They can grind through the repetitive diagnostic work that used to eat your sharpest hours. You still have to understand what they're doing and design the system they operate in. But the grunt work? Hand it off and go to sleep.
I'm running two businesses and my agent fleet handles work that would have required at least one more person a few years ago. That's not a flex. It's just the math now.
Part 4 — The Moat Is Still Time
None of this changes the basic math. The moat is time at the computer. AI makes each hour go further, but you still have to show up. You still have to be the person whose mind drifts to abstraction layers while everyone else is watching the game.
What's different now is the leverage. You can test ideas that would have taken weeks in a few days. You can maintain systems that would have needed a team. A solo developer in 2026 can carry a workload that would have been absurd in 2021.
But the leverage cuts both ways. Hermes is always running. There's always another issue filed, another optimization to chase. I've had to learn the hard way that the work is better when I walk away and come back the next morning. The architecture is cleaner when the problem has time to sit in the back of my head overnight. Rest isn't optional. It's part of the process.
You can learn how to productionize an application. Maybe the models and harnesses will eventually do it better. But for now, you can learn if you care. The moat for developers, with or without AI, has always boiled down to showing up. Now with agents on your device you can go further than ever. But take a balanced approach, because all-consuming computer interfacing is probably not a good thing.
Part 5 — Keep Going
If you're a developer reading this: don't get attached to your first working version. A functional prototype is not a product, and what you've already built should never stop you from building what you now know is better. The loops are shorter than they've ever been. The cost of starting over used to be measured in quarters. Now it's measured in weekends.
The only thing that hasn't changed is you still have to care enough to sit down and do the work.
Keep going.