Everyone's celebrating vibe coding. I'm watching where it falls short.

Cursor writes the function. The developer approves it. Ships it. Done in 90 seconds.

I’ve been spending a lot of time with software companies lately, and there’s a lot of talk about vibe coding. It’s results look impressive — until the function it creates lives inside a system with 12 years of accumulated decisions, undocumented dependencies, and three teams who each "own" a piece of it.

That's not a code problem. That's an institutional knowledge problem. And no amount of AI autocomplete closes that gap.

Here's what I keep seeing in enterprise software: the workflows that create real switching costs aren't the ones where code gets written. They're the ones where code gets understood — by auditors, by compliance teams, by the new VP of Engineering trying to map what they just inherited.

Vibe coding optimizes the fun part. The blank-page, greenfield, "let's build something" part.

Durable enterprise software lives in the other part.

The part with the change control board. The SOC 2 audit. The integration that touches payroll. The module nobody touches because the one person who understood it left in 2019.

And then there's the question nobody asks at he point of creation: who's going to support this?

Not deploy it. Not ship it. Support it. At 11pm when it breaks. Six months from now when the original developer has moved on and the AI context window that generated it is long gone. When the bug is in the interaction between the vibe-coded function and the legacy middleware it was never supposed to touch.

What developers are getting is real: faster prototyping, better autocomplete, a tireless rubber duck. For greenfield work, that's genuinely transformative. A skilled developer with AI is a different category of productive than a skilled developer without it.

But "better developer" and "fewer developers" are different claims. And "better at writing code" and "better at navigating institutional complexity" are different skills entirely.

The judgment calls — what to build, what to leave alone, who owns it when it breaks — those still need a human in the room. AI just makes the person in that room faster at the parts that were never really the bottleneck.