February 9, 2026 4 min read By Alex Dobrenko
So there’s this article from StrongDM about how they build software with AI that boils down to No Humans Allowed.
They call it “Grown code” which makes me think of Soylent for some reason, aka it’s dystopian and fascinating in equal measure, but most importantly: it’s maybe a sign of what’s to come for non-coding.
There’s a few ideas in here that are worth looking at with the opposite lens: what do I want to make sure I the human stay involved in:
“Why am I doing this?”
StrongDM calls this their “koan.” The implied follow-up: the model should be doing this instead.
My question would be sort of the opposite: Why is this task something I need to always do? And how do I get the model to do all the other boring stuff instead?
I don’t disagree with their framing! The only thing holding me back from doing more of this is probably… fear? Ego? The soft murmurs of my tiny human soul?
Scenarios, not tests
Their idea: don’t tell the AI what the code should do. Tell it what a real person needs to happen in real life.
Yeah this is probably the biggest unlock. Just talking to it like you’re talking to your friend. Be dumb, be clueless, be all of it. (I wonder if my own improv training is what makes it very easy for me to do this?)
Satisfaction, not pass/fail
They replaced “did it pass?” with “satisfaction,” basically: what percentage of people using this thing would be happy with it? Not binary. A spectrum.
I keep hearing this change from ‘deterministic’ to ‘probabilistic’ or, in english, from a binary Y/N to a ‘90% Y.’ That feels like the earth shattering thing that’s changing with all of this, in terms of how work is done and code (read: eventually everything) is written. You don’t need 100% success on everything, you need a good likelihood of success COUPLED WITH a strong “did it succeed for 100% sure” layer. Does that make sense? I’m asking myself mostly.
Deliberate naivete
StrongDM’s team actively tries to forget how things are “supposed” to work. Drop your inherited assumptions.
Yeah I agree lol. I feel like a little shit saying it, because I know how annoying it would sound if I were a software engineer hearing someone talk this way, but I do feel very much like NOT having baggage right now about how to do things is so so so helpful. Like I saw a tweet from someone about how they scaffold out the thinking of building something, a time zone selector, and I was like ‘why can’t you just… build it’? I know that the point this person was making was ’this is what you have to do because the code will get more complex and then you’re fucked,’ and they’re probably right and I will probably soon be fucked, but for now at least, at my level, it feels like the beginner’s mind of it all helps me move faster and do things.
Correctness compounds
Each good step makes the next step more likely to work. Same mechanic as the CLAUDE.md teaching loop.
Feedback loops ftw!! And not just for Claude but especially for you! The more you teach Claude stuff, the more you learn how to teach Claude stuff. That’s the real learning, and it’s why I’m wary about automating the feedback layer too much.
Use it more than feels comfortable
Their benchmark: $1,000/day on AI tokens per engineer. The point is the mindset. Most people are dramatically underusing AI.
I guess this is also true for me given that I’m on the $200 a month plan for Claude Max, a fact I would’ve thought IMPOSSIBLE five months ago. But yeah, the value I get from it is maybe 10x that? Idk. Maybe I’m deluding myself. As Claude Code said this morning, I “have 43 slash commands and can’t send an email. The system doesn’t matter. The schedule doesn’t matter. What matters is whether you can sit with the discomfort of being seen and do it anyway.”
The article summaries above were written by Claude. My reactions are my own. Read the full StrongDM article here.