Not Every App Should Become an Agent

One reason AI discussion gets tiring so quickly is that every useful idea gets universalized too fast.

Many applications moving toward agentic supervision does not mean every application should.

I may be in the minority here, but for many tools I depend on every day, I still want the old virtues:

determinism
inspectability
repeatability
direct manipulation
clear causal links between my action and the system’s state

I do not want my notes app trying to become my co-author. I do not want a personal tool “helping” me by nudging the artifact away from my intention. Sometimes I want software to obey, not infer.

That is not nostalgia.

It is a useful design boundary.

One mistake is to imagine the future as “everything becomes a prompt box.”

I think the real split will be cleaner than that.

Some products will become supervisory environments for agentic work. Those should lean hard into delegation, provenance, approvals, branching, and adaptive control surfaces.

Other products should remain deliberately deterministic and continue advancing the best of human-computer interaction without pretending that every workflow wants a probabilistic intermediary in the middle.

This matters because the rise of agents has created an odd temptation: to assume classical Information Architecture (IA), User Experience (UX), and UI research can now be skipped. As if a prompt box plus a foundation model absolves us from designing good software.

I think the opposite is true.

The age of agents raises the bar on interface design. In some products, we now have to design for supervision, evidence, reversibility, and trust. In others, we need to preserve clarity, control, and directness precisely because models are getting better at pretending they should sit everywhere. One may be surprised but there are still numerous apps where we do need a human in the loop, maybe validating, maybe signing off, maybe rejecting.

So the design question is not “where can I insert AI?”

It is:

where does agency genuinely help?
where does it create drag?
where must the system remain deterministic?
what kind of control does the user deserve here?

That last question is the one I care about most.

The strongest agent-first products probably will not be the ones that plaster prompt boxes, chat side rails and spray inference across every surface. They will be the ones that know exactly where to stop.

That, too, is product intelligence.

Previous: The prompt is not the product