Some of the most forward-leaning software teams I know now use tools like Claude Code, Cursor, or Codex to write the majority of their code. In some cases, 70–90%.

A few years ago, that fact alone would have been enough for many people to declare that we’d reached AGI. In hindsight, the goalposts moved, and probably correctly. Code generation turned out to be necessary, but not sufficient.

This post isn’t about AGI. It’s about something more mundane: shipping software.

What’s striking is that even on teams where AI writes most of the code, overall shipping velocity often looks… roughly the same. Output hasn’t exploded in the way you might expect.

When I talk to people about this, I usually hear one of two explanations.

The first is that it’s simply too early. We’re still in the first inning, and the real gains will show up once second-order effects compound.

The second is that code was never the bottleneck. Requirements, design decisions, coordination, review, testing — those were always the slow parts, so faster code generation doesn’t move the needle much.

I think there’s a third explanation that gets less attention.

I think we’ve accidentally created something that looks a lot like UBI — at least for software engineers.

We’ve seen something similar before. The internet didn’t show a clean, immediate boost to productivity metrics in Western economies. Instead, it made work more flexible, more distributed, and more optional. Email replaced mail. Zoom replaced travel. Remote work replaced offices. The efficiency gains were real, but they showed up as lifestyle changes rather than linearly higher output.

AI coding tools feel similar.

For many tasks, it really is an order-of-magnitude speedup to have an AI write the code instead of doing it by hand. Even if you discount the hype and call it a conservative 10×, that’s still enormous.

But that doesn’t mean most engineers want to think about 10× more features.

A highly motivated engineer might push themselves to ship dramatically more. A solid one might do somewhat more. But many will simply do a bit more than before — maybe 1.5× or 2× — and stop there.

If tooling makes you 10× faster, but you only take on 2× the workload, the remainder doesn’t vanish. It turns into slack. Less time pressure. Shorter days. More context switching tolerance. More room to breathe.

There’s another effect here that matters: these tools work immediately. An engineer who tries a modern coding agent for the first time often feels the speedup within hours, not months. The extra slack appears long before organizations adapt their expectations or incentives.

Until managers demand more, or markets force tighter competition, the default outcome isn’t radically higher output. It’s fewer hours of focused work to achieve the same results.

That’s what I mean by accidental UBI.

Not a check from the government, but a quiet redistribution of time. The work still gets done, but with less effort per unit output. Outside of elite tech bubbles, I suspect this effect will be even stronger.

I think this pattern will extend beyond software. Across knowledge work, large parts of the job will become easier, faster, and partially automated. For a while, that won’t show up as dramatically higher productivity. It will show up as people quietly wondering why their jobs feel lighter, and eventually, what they’re for.