Media Log
2027
2026
2025
AI-Assisted Engineering Is Addictive
Agentic development gives me the same sense of progression as a good video game. I keep wanting one more level.
An idea for a product used to come with a big question: how much work do I need to commit to before I could even have a working prototype? That time commitment alone would be enough to put me off. There’s only so much time you have, and realistically you can’t play ten 100-hour games in just a month.
With agents, I can now get to something working much sooner. I can learn as I build and experiment with something in front of me, trying to improve it along the way. Ideas that once felt like too much of a commitment now feel worth a go.
It also feels a lot like playing video games with a strong sense of progression. Every new project is a new game, and you level up fast at the beginning. As the codebase grows, progress still slows down, but it no longer feels like grinding from level 59 to 60 in WoW.
And the next project you start is akin to starting another character on the same account. The gold and items you’ve collected carry over. Here, that’s the skills and knowledge from the last build: what to ask for, what to watch out for, and what you can reuse. You get through the early levels even faster with the help of agents.
I’ve been finding this style of development so addictive. I sit down to try one thing, make some progress, and end up staying longer than I planned. I guess the FOMO about leaving some of my weekly limit unused also adds to it.
The effort axis
AI compresses wide, shallow build work. Roadmaps should size design-heavy work apart from the cost of getting a feature wrong.
Three familiar prioritization methods make room for effort:
- RICE divides reach × impact × confidence by the whole team’s effort.
- Value versus effort puts effort on one axis and expected value on the other.
- MoSCoW sorts work by necessity, then checks whether the Must Haves fit the team’s capacity.
With agentic engineering, the effort dimension has been compressed, an L might not be that large anymore. We could perhaps break effort into two types. One is complexity, work that is heavy on design and the architecture. The other is a large in surface but simple or shallow. For the large but shallow kind, an L or even an XL might just take a day or two for agents. Complex, architecture-dependent tasks are those that still need people to oversee.
Prioritization should then value effort size differentlty by type. A uniform cut to every RICE effort estimate would leave the ranking outdated. The actual estimation should be based on whether the type of effort has human cost or not. If task is simple, clearly specifiable, light on architecture, low blast radius or the work is off the critical path, a team can hand the whole attempt to an agent and ship it. The old L or XL label is no longer as daunting.
How to Set Goals Before Product-Market Fit
Before product-market fit, set goals on what you learn, not on numbers you can't move yet.
Before product-market fit, a revenue target is a guess wearing a number. The goal that holds up is one you control this week, like how many users you spoke to, or how many changes you shipped because of what they said.
The failure they keep coming back to is the team that sets a goal it can’t move, misses it, and learns nothing from missing it. A goal should tell you what to do on Monday.
What I’m taking from it, until outcomes are predictable, goals track learning. After that, they track outcomes.
Handmade code
Factories didn't make handmade clothing extinct. Some exquisite designer clothes are still made by hand. Maybe code will be the same.
AI code shouldn’t be slop, but it doesn’t need to beat the best engineers. If 500 software companies had equally sized teams, the top 1% of their engineers could staff only five; the other 495 still ship software people use. AI only needs to clear that bar.
Food already works this way. A frozen lasagna may not beat a chef’s, but it might beat yours. It’s cheap, fast, and good enough. People still cook when the meal matters and reach for the freezer when it doesn’t.
Software could go the same way. Payments, auth, and anything handling money or data deserve careful design. CRUD, glue, and another form component just need to work. Subpar is fine there, and AI code may not be subpar anymore.
Handmade code won’t go extinct, it’ll stay in the parts of a system where failure costs the most.