The agent demos are selling the wrong miracle
You’ve seen the demo. Someone types “book me a flight to Lisbon and a hotel near the old town,” and the agent clicks through three sites, fills the forms, and confirms a booking. The room claps. The clip gets two million views.
Now ask yourself the only question that matters: would you actually let it press “confirm” on a non-refundable flight?
Of course not. So what was the demo for?
The trick is in what they don’t show
The reason these clips are exciting is the same reason they’re useless. They show an agent doing the one thing you’d never delegate without watching: an irreversible action, unsupervised, on your behalf. It books, it buys, it sends, it posts. The thrill is the absence of a human. That absence is exactly what makes it unusable for anything you care about.
Watch what gets cut from the demo. Nobody films the part where the agent picks the wrong Lisbon hotel because there are four with similar names. Nobody films the refund call. Nobody films you explaining to your boss why a contract went out with last quarter’s numbers in it. The demo ends at “confirmed” because everything interesting happens after.
We’ve built our whole industry’s hype around the least useful 5% of what an agent does, because it’s the part that looks like magic on video.
The boring 95% is the actual product
Here’s what an agent is genuinely good at, none of which makes a good clip.
It reads twelve hundred words of a thread you were cc’d on and tells you the one sentence that needs a decision. It checks whether the number on the dashboard moved for a real reason or because a tag broke. It notices that the competitor quietly changed their pricing page and tells you what changed. It pulls the three documents you’ll need for your 10am into one place so you stop hunting through Drive at 9:58.
This is reading, synthesis, and judgment about what’s worth your attention. It is enormously valuable and completely unfilmable. It saves you the hour you didn’t know you were spending. And it carries almost no risk, because the worst case is a summary you disagree with, not a flight you can’t cancel.
Acting is fine. Acting alone is the problem.
None of this means agents shouldn’t do things. They should. An agent that finds the insight and then drafts the LinkedIn post about it is more useful than one that just hands you the insight.
The fix isn’t to ban action. It’s to put a human at the one moment that’s irreversible. The agent does the reading, the thinking, the drafting. Then it stops and shows you. You approve, or you don’t. That single step is the difference between a tool you can leave running and a liability with a cron job.
It also happens to be the step every demo skips, because “and then she read it and clicked approve” is not going viral.
What to actually look for
Next time you’re evaluating one of these things, ignore the autonomy. Ask what it does on a Tuesday when nobody’s watching and nothing’s on fire. Ask what it reads, what it notices, and what it hands you. Ask where the human sits.
The product worth paying for is not the one that does the most without you. It’s the one you’d trust to run every morning while you’re asleep. Those are very different machines, and only one of them is in the demo.