I Went to Ship 26. Everyone's Shipping Agents. Nobody's Shipping the One Thing That Matters.
I spent the afternoon/evening at Vercel's Ship 26 in Berlin — a dark auditorium, a couple hundred developers, and a row of slides everyone in the room was photographing at the same time. The whole event is one message, repeated from a dozen angles: software is agentic now. Agents deploy. Agents build agents. Agents watch production and roll themselves back.
And the numbers back it up. Vercel said over half of all deployments on their platform now come from coding agents — up from under 3% six months ago, a 17x jump. AI Gateway traffic went from 2 trillion to 20 trillion tokens a month. One talk put the same point a different way: ~60% of enterprise merged commits are now written by agents. This isn't a forecast. It's already the baseline.
I came away convinced of the trend. I also came away convinced almost everyone is reading it backwards.
The slide I can't stop thinking about
It wasn't on the keynote stage. It was in one of the engineering talks, and it was an honest one. A team put up their own internal data on rolling out AI coding tools. The capability line went straight up — exactly what you'd expect.
The results line didn't.
It rose a little, then flattened out and stalled around a 40% productivity gain. They'd labeled the curve themselves, on their own slide, in big letters: "THE PLATEAU." Next to it, a faint dotted line showed where they'd be if it scaled linearly — way up and to the right. The gap between those two lines — capability racing ahead, outcomes flattening — is, more or less, the entire business I'm building.
It's the same shape as the METR study from last year: experienced developers were given AI tools, expected to go faster, and measured themselves 19% slower. More capability did not become more shipped. It rarely does on its own.
The most human thing said all day
Here's the line from the keynote that stuck with me. Vercel's CEO, talking about all these agents companies are rushing to build:
"Agents are free. Free as in free puppies. Everybody loves puppies. But they pee on your floor. They eat your furniture… agents are software, and we all know that software is never done. Someone has to maintain them."
Read that again at an event whose entire theme is autonomy. The headline is "agents do the work now." The fine print, said out loud on stage, is: someone still has to finish and own the thing.
And it wasn't a one-off. Their whole agent stack is built around a human gate. The Workflow SDK pauses when the job needs human input. Their production agent will investigate an incident on its own — and then "anything that changes production state waits for a human to approve it." Another talk reframed the entire job of an engineer: the question, the slide said, is no longer "how do we help engineers code faster?" — it's "what human-review artifacts do we still need, and how do agents prepare them, so humans operate at the right checkpoints?"
The most advanced agent tooling in the room is architected around the assumption that a human stays in the loop. That's not a footnote. That's the actual shape of the thing.
Here's the part that doesn't sell infrastructure.
When agents write half your code, building stops being the hard part. And a bottleneck doesn't disappear when you remove it — it moves up a layer. It moves to the part agents can't do for you: finishing. Owning it. Showing up on Day 4 when the novelty's gone, the deploy is boring, and nobody's watching.
The faster building gets, the more half-done projects pile up. The graveyard of "90% done, runs on my machine, never shipped" is bigger this year than last — not smaller. More capability, same completion gap. That plateau slide was about a funded engineering org with accountability built in. Now picture a solo dev, alone, on a side project, with none of it.
A conference full of developers who can now build anything in a weekend is, quietly, a room full of people who are about to have more unfinished projects than ever.
Why "the build is easy now" is the trap, not the win
The thing nobody's shipping
So here's the honest takeaway from an evening surrounded by autonomous everything:
Vercel keeps a human in the loop on production. Almost nobody keeps a human in the loop on themselves.
You'll wire up approval gates, scoped permissions, and rollback safety for your agents. Then you'll start a side project alone, with zero gates, no one expecting the next step, and act surprised when it dies at 85% like the last three did.
AI tracks. A human reads. A dashboard shows you activity; it doesn't notice your absence. The accountability layer — the human who reads and names the move you're avoiding — is the one part of this whole stack that nobody can deploy for you. And it's the part that decides whether anything you build with all this new power actually ships.
The honest next step
If your projects keep dying at the same point — usually your point — more agents won't fix it. Knowing where your loop breaks might.
I built a free diagnostic for exactly that. Seven questions, two minutes, no signup to see the result. It names where you stop and the one move that breaks it.
Everyone left Ship 26 ready to build faster. The harder, more valuable question is whether you'll finish. Find out where your loop breaks: mvpbuilder.io/ship-readiness
*Building in public. Day 131
Top comments (0)