Terminals are the best interface for LLMs, not for humans

Terminals are the best interface for LLMs, not for humans.

I didn't think this take was particularly controversial, but of all the things in my Cyboflow launch it got the most heat.

First, before the pitchforks come out I'm not coming for your TUI. If you had a childhood fascination with the Matrix and want to spend your days working out of it then it can remain in your warm, alive hands. But I don't think that's true for most people and that's core to why I built Cyboflow.

But first, why are we seeing this explosion of TUIs in the first place? To me, the reason for that is simple. It's the ideal home for an LLM. Terminals are an extremely powerful text-only interface and that's the perfect match for a language model.

They're generally not the right ones for humans though. There's a reason GUIs replaced DOS and most developers would work out of an IDE and not their terminal. Text by itself conveys a lot less information than a rich graphical interface.

Consider these two airline tickets, which one is easier to read?

The information in both is text, but the designed version is much more scannable.

That's not even getting into the power of images and graphs, but even when you're just dealing with text having a visual dimension makes it much easier to digest.

I'm pretty explicit that Cyboflow is not an IDE and it's not a TUI. IDEs are going to become a relic of the past and TUIs are a stepping stone.

IDEs are built around writing code, when you open one the center of your screen is the code you're working on.

In a TUI, the terminal is central. Increasingly, that means a bunch of terminals running agents, but still the primary focus is the terminal itself.

But terminals are fundamentally constraining homes.

You have to fit an illustration of a UI concept into ASCII art. You have to build a separate rich text composer just to move your cursor with a mouse. You're either providing an all too brief summary or a wall of text because you can't include something as simple as a toggle.

I call Cyboflow an agentic development environment because its point is to prioritize the thing you're trying to focus on. Increasingly that means your agents, not your code, and not your terminal.

I don't have anything against terminals, and in fact every Cyboflow session has least two of them. They're still critical to any real development work. They're just not the place you should be spending all your time.

As agents get more capable and you're managing more of them, you need higher bandwidth interfaces to manage them that help you quickly scan the information you need and provide the right input. The only way to do that is to add a dimension that goes beyond text.