Are customers building their own tools now?

In July 2025, the research organization METR published a study meant to give a clean answer to one very simple question. Does AI make experienced developers faster? Sixteen experienced open source developers worked through 246 real tasks in their own projects, and a random draw decided which of those tasks they were allowed to use AI for and which they were not.

The result was a proper reality check. Beforehand, the participants expected AI to make them 24 percent faster. Even after the work was done, they still felt 20 percent faster. The stopwatch disagreed. In reality, they were 19 percent slower.^[1]

What happened next is even more interesting.

When METR tried to repeat the experiment just a few months later, it produced no usable result at all. The reason is spelled out drily in the methodology notes. They simply could not find enough developers willing to work without AI anymore. Not even 50 dollars an hour changed that. And some of those who did take part never even submitted up to 50 percent of their tasks, because they refused to do them without AI.^[2]

That proves a dependency more clearly than any survey could. People left real dollars on the table just to avoid giving up their AI assistant.

And that reflex reaches well beyond programmers by now. The barrier to building your own software has dropped so far that companies simply click their tools together themselves. Work they would once have paid an agency for without a second thought, they now do in-house. That was exactly the topic of a recent episode of our podcast Buzzsozial.

Are agencies dying now?

Our episode on Agentic Coding featured Fabian Rossbacher. He founded the SEO Days and has been a developer since the nineties. Today he builds software on his own at a pace he puts at 25,000 to 35,000 lines of code a day.

At some point during the conversation, he said this:

“Customers have got smarter, because they build their own tools now. The death of the agency has begun, and it is already well under way.”^[3]

That is not a forecast, it is an observation, and the first half of it is hard to argue with. Fabian supplies the best example himself. A year and a half ago, a friend of his was still building classic WordPress sites with her small agency. For the past nine months she has worked only in Cursor, client projects included. An agency owner has effectively replaced her own way of producing work.

The speed of it is remarkable too. Two years ago, a customer would have asked for a quote for a simple internal tool. Today they just build it themselves. So Fabian is not far off with his conclusion. Whatever a customer can pull off quickly on their own is barely sellable anymore. The quick dashboard, the standard report, the small tool that mainly exists so that one exists.

But none of that has much to do with end-to-end tool development or content production as such. Those have not disappeared. They just demand something different today than they used to.

The tunnel vision of vibe coding

The METR participants felt faster with AI. They were not. Anyone who has just built their first tool probably knows the feeling. It runs, it looks tidy, and after a few hours you have something that used to take weeks. What you do not see in that moment is everything that only shows from the outside.

You cannot see whether it is even the right tool, for a start. Once code is no longer the barrier, the only bottleneck left is the idea itself. And where that is missing, people build whatever is closest to hand. In Fabian’s circle of founders there is a running joke about it. When someone says the company really has to do something with AI now, the answer comes back: “Sure, build another dashboard. Maybe another dashboard will solve the problem.”

This probably sounds familiar. Building is suddenly the easiest thing in the world. But the instinct for what actually solves a problem is not something AI ships with.

It is just as hard to judge from the inside whether the result is actually any good. When you built it yourself, you mostly see what you wanted to see, and that goes for professionals as much as for beginners. There is a reason why even seasoned developers have strangers test their interfaces. Anyone building their first tool usually has no one like that, and so gets no honest feedback.

And just because something runs today does not mean it will still work in six months. If you simply tell AI what you want, you get a slightly different route through the code every time. That usually only surfaces later, when you want to change one thing and something breaks somewhere else entirely.

So anyone developing seriously with AI has to give it firm rules up front. A definition of how to build, so that it follows the same principle every time. That is the big difference from what most people picture when they hear vibe coding. Rules like that do not appear in a day. It often takes weeks before AI reliably works the way you need it to, and you throw away a lot of what it built along the way. None of that work is visible in a finished tool.

And as soon as what you built is meant to reach customers instead of only running internally, something else comes into play. Whether anyone finds it at all is the thing you can see least from the inside. That applies above all to websites, which some companies now build themselves just like their tools. A site like that goes online quickly, but that alone does not get it into Google or into an AI answer. What content it needs, how a page has to be structured so that search engines and language models understand it, and who links to it from outside, is not decided by a Cursor subscription. And when everybody can build a site in no time, there are masses of them, and the individual one stands out even less.

And the agencies?

So are they dying?

The ones in real trouble are those whose work consisted mainly of what a customer now builds themselves. Nobody frees up budget for that anymore.

But none of what you cannot see from the inside can simply be built alongside. Whether something is the right thing, whether it is any good, whether it holds up and whether anybody finds it, you can only tell from the outside. And the more that gets built, the more that view matters. On our end, too, it shifts what customers need us for.

The METR participants were sure they had got faster. They only found out that they had not because somebody measured from the outside. That outside view is exactly what a good agency delivers, and it is the one thing no customer can build for themselves.

Sources