Explaining AI agents is like explaining 3D printers

Have you ever tried explaining why somebody needs a 3D printer?

"It prints stuff" sounds profoundly unimpressive.

3D-printer owners make it worse by enthusiastically showing you the ugliest plastic objects ever produced by mankind. A bracket. A spacer. All covered in those little horizontal print lines and usually made from whatever radioactive-green filament happened to be loaded.

And yet the owner is delighted.

The useful part is almost impossible to communicate. Explaining it abstractly sounds underwhelming: "You can make small plastic objects."

Then you get a printer yourself, something breaks and suddenly you think: hold up, I could make a replacement for that. Then you print a hook. Then a cable guide. Then a jig for drilling two holes at exactly the right distance. Then a piping joint nobody would deliver on a Christmas Eve when you need to fix that emergency leak. Then you fix your daughter's broken toy in 20 minutes - and you both have fun in the process.

Soon you start seeing printable problems everywhere. You might even get to the "having a hammer - everything becomes a nail" moment.

The important thing about 3D printers is not what they can print. It's what you start noticing after you've used it once

Explaining AI agents to non-tech people is similar

"It can use tools and do tasks for you" is about as convincing as "it prints plastic thingies." But once an agent handles one real annoyance, you start seeing delegatable problems everywhere.

A recent example: I messed up some paperwork and had to re-pass a motorcycle theory test for my driving license. The local traffic police website offered 30-question demos using real exam questions, but no downloadable question bank. So I pointed an agent at it. Asked it to inspect the site and its backend requests, write a scraping script and pull about 550 questions to JSON. Then I had it turn those into a tiny mobile app, practiced with it for a week and passed the test perfectly.

It sounds thoroughly unimpressive. "I made a driving-test app." Fine. There are thousands of those. Except this one contained the exact questions I needed, from the exact awkward website I had to use. It solved one problem, for one person, probably once. It's the same mental shift, I spotted a "printable problem".

Economy of one

The original futuristic pitch for 3D printing was that we would manufacture everything at home. Plates, shoes, furniture, possibly a new kidney before breakfast.

That mostly missed the point. Mass production is extremely good at mass production. IKEA is not worried about my little plastic machine.

A 3D printer is great at the opposite: economically irrational customization. It can produce one object so absurdly specific that designing, packaging, stocking and shipping it as a normal product would make no sense.

Agents seem to be good at the software equivalent: economically irrational automation.

Traditional automation needs scale. If a process happens a hundred thousand times, somebody builds software for it. If it happens twelve times a year, requires opening two websites and a spreadsheet, and only makes sense inside your company, everyone agrees to keep doing it manually until the end of time.

Agents lower that threshold to almost zero. They still get confused, click the wrong thing and occasionally announce success while staring directly at an error message. You should not let one freestyle inside your banking app because some dude on Twitter called it an "AI employee".

But small, supervised, low-risk jobs are different. The good applications aren't "launch a business while I sleep." They are "check these invoices against our weird internal rule," "visit these twelve pages and tell me what changed," or "research this new tax deduction law and tell if we're eligible." Or even "will my car fit into the 10:30 ferry boat". I used to awkwardly siri-dictate reminders for future me, now I just tell Claude what to do.

These tasks live in the enormous long tail that was never worth turning into software. That is exactly where agents get interesting. Just like 3D printers.

So no, I still can't explain to you why you'd want either one. That's kinda the whole point. Nobody managed to explain it to me either.