I can generate a year of work in an afternoon now. I have learned to be suspicious of how good that feels.
The feeling is real and it is intoxicating. I describe a system, and an agent returns the schema, the migrations, the handlers, the tests, the docs, the little README nobody used to want to write. What would have been a quarter is now a long lunch. I push back from the desk and there it is — gleaming, complete, more output than I produced in some entire years of my career. And every time, somewhere underneath the satisfaction, a quieter instinct asks the same question. Is this the good stuff, or is it just cold and fizzy?
I have started thinking about it as the difference between soda and wine.
Two kinds of knowledge
Soda is manufactured. You combine the inputs, apply pressure, and it is ready before you've finished reaching for a glass. It is genuinely good — cold, sharp, satisfying on a hot day, and there is nothing fraudulent about it. The agents are extraordinary soda machines. Ask, and the can is in your hand, beaded with condensation, indistinguishable at the first sip from something that took far longer to make.
Wine is the other thing. You cannot pressurize wine into existence. It requires a particular grape in a particular soil, a season that went the way it went, and then years in the dark doing nothing you can watch. The value is in the time. Rush it and you don't get young wine — you get grape juice, which is a fine drink and an entirely different one. No machine has ever shortened that clock, because the clock isn't an inefficiency in the process. The clock is the process.
Code, it turns out, has both kinds in it, and we spent decades unable to tell them apart because both were slow. Writing the handler and knowing why the handler should exist were welded together, both expensive, both yours. So we priced them the same. We paid engineers for output and quietly received the judgment for free, riding along inside the typing.
The machine just unwelded them. The soda — the schema, the boilerplate, the third CRUD endpoint that looks like the first two — pours instantly now and costs almost nothing. What it cannot pour is the wine: the knowledge of which schema is wrong in month nine, why this customer's "simple" request will quietly metastasize, where the bodies are buried in a system you have lived inside for three years. That knowledge has only ever come from time, and time is the one input the machine can't fake.
Why the fast stuff fools you
Here is the trap, and I have walked straight into it. When output is free, you produce a great deal of it, and volume reads as progress. The afternoon's gleaming pile looks like a year of work because it is the size of a year of work. But size was never the thing. A year of work used to be valuable partly because surviving it taught you something — you came out the other side knowing the domain in your hands and your gut, not just your notes.
Generate the same artifact in an afternoon and you skip the tuition. You hold the output without the understanding that used to be inseparable from making it. The first time a system I'd "built" in a day broke in a way I couldn't immediately read, I understood the bill had simply been deferred. I had the soda. I had never made the wine. And debugging is a wine problem — it asks what you know, not what you can produce.
This is the quiet inversion of 2026 that I don't think we've fully metabolized. For my whole career the constraint was making the thing. Now making the thing is the cheap part, and the binding constraint is the slow, contextual, lived-in knowledge that tells you which thing to make and why — the part that still moves at the speed of a season, because it forms by living with a problem and cannot be hurried by any amount of compute.
What this changes
At my company we ended up writing it down as a stack, because once you see it you start ranking everything by it. Quality over coverage. Coverage over speed. Speed, the thing the whole industry is selling, sits near the bottom — not because it's worthless, but because it's the one part that's now abundant, and you don't protect what's abundant. You protect what's scarce. What's scarce is the deep, provenance-carrying knowledge: not just the answer but the trail of why, the context that makes the answer trustworthy six months later when nobody remembers the conversation.
So the discipline reverses. Instead of asking "how fast can we ship," the better question became "what do we actually understand here, and how do we keep it." The agent drafts; a person decides whether the draft reflects something real or just something plausible. Plausible is what soda is best at. Plausible is exactly the failure mode of a machine that can produce the shape of expertise without any of the time underneath it.
I think the teams that win the next few years won't be the ones generating the most. Everyone can generate now; that race is over and the prize is a participation trophy. The edge moves to whoever cultivates the slow knowledge fastest — who lives closest to real problems, retains the hard-won context instead of letting it evaporate, and uses the abundance of cheap output to buy time for the expensive understanding. Let the machine pour the soda all day. Spend the hours it frees on the cellar.
I still get the rush. I'd be lying if I pretended the afternoon-that-makes-a-year doesn't thrill me a little every time. But I've stopped mistaking the volume for the value. The output was never the scarce thing — it just looked scarce because it used to be slow. The scarce thing is the wine, and the only way to make wine is still, stubbornly, to give it the years.