From Answers to Actions: How AI Agents Are Changing What We Expect Computers to Do

Every earlier generation of software waited for you to click.

Agents don't wait. Give one a goal and it breaks the goal into steps, then starts doing them — booking, filing, editing, sending — without pausing for you to approve each move.

The Undo Button Doesn't Scale

Answers are cheap to be wrong about. You read one, notice it's off, ask again. An action isn't like that. Once something's been sent, moved, or paid, there's no gentle "let me try that again."

That's the real shift, and it's easy to miss because the model itself didn't get smarter overnight. What changed is that we removed the checkpoint. Decades of software assumed a human reviews the output before anything happens in the world. Agents are built to skip that step by design.

We've Stopped Grading Answers. We're Grading Judgment.

A chatbot gets evaluated on whether its response was correct. An agent gets evaluated on whether it knew when not to act — when to pause, ask a clarifying question, or stop entirely.

That's a harder problem than generating good text. The interesting engineering work has moved from "how do we get a better answer" to "how do we build something that understands the size of its own mistake before it makes it."

What This Means for the Software We Build

Every product adding agents now has to answer a question chat interfaces never had to: when it's wrong, who finds out first — the user, or the consequence?

The interesting products right now aren't the ones with the smartest agent. They're the ones that decided, deliberately, what that agent is and isn't allowed to touch.