Software Was Never Really For You

Karpathy posted a thing about CLIs being the sleeper format of the agent era - old, text-based, composable, and it turns out that's exactly the shape of tool an LLM agent can grip. Install a Polymarket CLI, install the GitHub CLI, point an agent at both, and watch it build you a terminal dashboard in three minutes while you get coffee. He asked the right product question too: can agents access and use what you've built? Is your docs exportable as markdown? Do you have a CLI? An MCP endpoint?

He is completely right. I've been living here for a while.

But the more interesting thing - the thing underneath the tweet - is what this implies about software in general.

Software is not for humans to use anymore.

That sounds wrong. It isn't. You use an agent. The agent uses many agents. Those agents use many tools, CLIs, APIs, code, data. And at the end of that chain, what arrives in front of you is exactly what you needed to see - already processed, already synthesized, already shaped to your actual question.

This is exactly how vision works. Your eyes don't hand you a stream of photons. Your brain doesn't show you raw wavelength data. Something in between does an enormous amount of invisible work, and what lands in your consciousness is just: the thing. The scene. Already interpreted.

We are building the visual cortex for software. The raw interface layer - the CLIs, the APIs, the legacy docs written for humans who no longer need to read them - that's the retina. You'll never see it. The agent will.

The question isn't "is your product usable." It's "is your product legible to something that isn't a person." Different question entirely. Most products fail it.