Every AI tool I open lately greets me the same way. A blinking cursor in a text field, waiting for me to type. Claude Code, Codex, all of them. The most powerful models we've ever shipped, and the interface they wear is the same one we used to talk to IRC bots in 1999.
I get why. Chat is the lazy default. It's the cheapest interface to build, it ships fast, and it makes a demo look magical. But I've started to think chat is a placeholder we mistook for the destination. The real endgame isn't a smarter chat box. It's no chat box at all.
Chat makes you do the hard part
Here's the thing nobody says out loud. A chat interface quietly hands the entire job back to you.
You have to know what to ask. You have to phrase it well. You have to read a wall of prose that comes back, hold it in your head, and translate it into action yourself. The model did the thinking, but you're still doing the interface work, in your own working memory, for free.
This gets especially silly with a coding agent. The work I'm actually doing is spatial and structural. Files, diffs, dependency graphs, call stacks. None of that is a paragraph. Yet we keep squeezing it through a one dimensional stream of text, like reading a map by having someone describe it to you over the phone.
When my agent refactors a function and then tells me, in three sentences, what it changed, that's not help. That's homework.
What the alternative looks like
So flip it. The model can already generate anything. Code, markup, layout, components. Why does it keep generating a description of the work instead of generating the interface to the work?
Picture it rendering the controls the moment actually needs.
It refactored something? Give me a diff view with accept and reject toggles per hunk, not a summary I have to trust. Let me see the change and act on it in the same surface.
It's debugging? Surface the call graph and let me poke it. Let me click the frame I care about instead of asking "what called this function" in English and waiting for the answer to scroll by.
It's about to touch twelve files? Show me those twelve files as a checklist I can scope down before it runs, not a confession after.
The interface stops being a fixed shell wrapped around the model. It becomes a fluid output of the model. The UI gets generated per task, per moment, shaped to the exact decision I need to make right now.
This is really about trust
The deeper reason I care about this isn't aesthetics. It's trust.
Chat hides the agent behind prose. "I refactored the auth flow and updated the tests." Cool. Did you though? I can't see it. I either take the sentence on faith or go dig through the files myself, which defeats the entire point of having an agent.
A generated interface makes the work inspectable by default. The diff is right there. The plan is right there, as steps I can edit before they execute. The agent stops saying "trust me, I did the thing" and starts saying "here's the thing, shaped so you can verify it in two seconds."
That shift, from describing the work to presenting the work, is the whole game. An agent I can verify at a glance is an agent I'll actually let off the leash. One that only narrates is one I'll micromanage forever.
Chat doesn't die, it gets demoted
I want to be honest about the counterpoint, because there's a real one.
Chat survives for a reason. It's universal. It handles ambiguity no button ever anticipated. Sometimes the thing I want to say has never been a control on any screen, and language is the only tool flexible enough to express it. Kill chat entirely and you lose that.
So I don't think the answer is no language. I think it's language as the entry point, generated UI as the response.
I type what I want in plain words, because that's the most expressive way in. But what comes back isn't more words. It's the right interface for what I just asked, assembled on the spot. Language in, interface out. Chat becomes the front door, not the whole house.
The part I keep coming back to
We spent years arguing about which chat product is best. Better memory, longer context, faster responses. All real, all incremental, all still inside the same little text box.
The jump I'm waiting for isn't a better answer in the box. It's the box dissolving, and the model handing me exactly the surface I need to make the next decision and move on.
The agents are already smart enough to do this. We just haven't asked them to stop talking and start building the thing I'm actually looking at.
That, to me, is the endgame.
Top comments (0)