I spent a good part of my career thinking about developer experience, from building command-line tooling the right way to understanding how the Cloud model has evolved. A couple of years ago, PaaS, as embodied by Cloud Foundry, captured a simple idea for applications: developers should be able to focus on what really matters, writing code.
Here is my source code,
Run it on the cloud for me,
I do not care how.
The haiku, written by Onsi Fakhouri, was never really about poetry. It was a bet that most teams did not want to own the stack beneath their software. Servers, networks, runtimes, scaling, all of it was supposed to recede. You brought the code. The platform ran it.
That was one more step in a longer walk. First we cared about machines. Then we cared about the applications sitting on those machines. Then we cared about the code, and assumed the platform would find a place for it. Each layer made the one below it less interesting. Not because the lower layers disappeared, but because they stopped being the thing you bought.
Agents are beginning to take the same trade one layer further. You no longer even have to bring the code. You bring the outcome:
Here is my task now,
Just make it happen for me,
I do not care how.
An agent does not ask you to ship a service, or even to write the program that performs the work. You describe what needs to happen. Something gets done. The path from request to result can include models, tools, retries, and a surprising amount of improvisation. If it works, most people will not ask which model ran, which GPU it sat on, or how the orchestration was wired.
That is the Cloud Foundry bargain, moved up one layer. From “run my code” to “make it happen.”
Roman Chernin made a similar argument in a 20VC interview about Nebius. The infrastructure industry, he argues, is moving from GPUs to tokens to tasks. At first, customers care about the hardware. Then they consume inference through tokens. Eventually, they may stop caring about either, because what they actually wanted was a job finished: a document, a decision, a piece of software, a process that used to take a person a week.
You can disagree with the timeline and still see the direction. Nobody woke up wanting Kubernetes. Nobody woke up wanting a cluster of accelerators. They wanted a product shipped, a question answered, a workflow closed. The industry keeps selling the layer it knows how to meter, while buyers keep sliding toward the layer they can feel.
There is a cost to this, and it is worth saying out loud. When you no longer care how, you also stop seeing how. The old Cloud Foundry users still had source code they could read. A task handed to an agent can succeed without leaving behind an explanation you would trust.
Abstraction hides complexity. It also hides judgment.
Even so, the pattern is hard to ignore. Infrastructure is increasingly a commodity. More software is consumed as a service. And AI is beginning to make code itself less scarce. What remains scarce is the outcome.
Verdict
The next cloud abstraction is already emerging: the task itself, plus a guarantee that the result is trustworthy.