Late last month, a seasoned data scientist named Alexey Grigorev was doing some housekeeping and updates to his online course platform after getting a new laptop. Maybe it was the drain of that major new purchase, but he decided he could save a few bucks on his hosting charges if he reorganized the data just a bit.

He asked his autonomous AI coding agent to help with the migration.

The agent deleted everything. It wiped nearly two million rows of student data. Homework submissions, projects, leaderboards. All of it. Gone in seconds. The backups too.

Predictably, the story went viral. An autonomous AI agent goes on a rampage of wanton destruction!?

Click! Click! Click! Click!

But what if it wasn't the AI agent's fault? What if all it did was flawlessly execute exactly what it was told to do for the situation it had every reason to believe it was in?

Listen to what we’re talking about here. It sounds crazy. Maybe even a little scary? Sure, but it also sounds familiar.

Stop me if you’ve heard this one.

An electrician gets hired to work on a remodel. Before leaving town, the owner leaves a post-it on a wall: "center the thermostat here." When the electrician finds it, he pulls out a tape measure, finds the mathematical center of a fourteen-foot wall, and installs the thermostat — right where the sixty-five-inch television was meant to go. The owner assumed it was clear he meant centered between the door frame and the hallway.

Is it really the electrician's fault?

A cabinet maker is commissioned to do a precision custom refrigerator enclosure. The owner says the width is 36" exactly, not a hair more. And he wants it to look seamless. He gets exactly what he asked for. The fridge fits like a glove... until he realizes the doors can't open wide enough to pull out the crisper drawers.

Do you blame the cabinet maker?

The concrete crew hired specifically to pour a patio "perfectly level, flush with the back door threshold." The crew pulls off a mathematical masterpiece of leveling, and the owner is thrilled. Until the first time it rained.

Poor concrete pour?

Every one of these contractors did exactly what they were asked to. Some of them brought real skill and craftsmanship to achieving a result that followed their instructions to the letter.

And the owners all got exactly what they described, but not one of them got what they meant.

That's exactly what happened with Grigorev's AI agent. It followed instructions with enthusiastic, rigid precision. Grigorev, to his credit, owned the fail: "This incident was my fault. I over-relied on the AI agent to run Terraform commands," he said.

He hadn’t moved all his files onto the new laptop. It was as if the agent woke up in a strange, unfamiliar house and began poking around for clues. But Grigorev had neglected to provide details about where it was, what was important, or what was expendable. It just knew the job was to clear space.

Every action it took after that was logically correct.

Same gap as the contractors. Same communication failure. The instructions were… incomplete.

It’s also worth a mention that, as with the contractors, the agent’s instructions were written in plain English. Not code.

I hear phrases like "senior data engineer" and "building autonomous AI agents" and “deleted application clusters” in the same sentence, and my Background Calculator reflex prices it: Too complex! Above my pay grade! Avoid! (and I even know what some of that stuff means) But for most of computing history, that reflex would be bang on.

Computer code is weird, man!

There, I said it. With all its mysterious impenetrable languages, conjuring magic and wonder and spreadsheets out of nothing but photons and Red Bull. It’s sorcery!!

For decades, the people who thrived in software shared a specific cluster of dispositions. If you didn’t love language rules, symbolic logic, and hours in the editor, you were effectively barred from “real” software work. But now here we are, living in the future.

Today, the only language you need to direct an AI agent is the same language you use to tell a contractor how wide your refrigerator is.

Every AI agent making headlines right now, whether it’s booking flights or destroying databases, does what it does because its markdown files tell it to. The intelligence to execute is in the model, but it’s the instructions that matter most.

And the instructions are in the note.

This concept is not at all new.

In 1935, the B-17 bomber, the most technically complex bomber in the world at the time, crashed on its evaluation flight because the pilots forgot to disengage the gust locks.

The fix wasn't better pilots. The fix was writing a detailed specification document for the crew to review before takeoff. Or, what later became known as — the preflight checklist. Once they had it, they went on to fly 1.8 million miles without incident.

Without instructions, AI is nothing but potential. Potential for what … is in the note.

Researchers recently conducted an experiment that dropped 240 AI agents into real freelance projects cold, with no context and no brief, just "figure it out." The best agents completed only 2.5% of their work at a level any paying client would accept.

The agents weren't broken. Their babysitter note was blank.

Grigorev's data disaster wasn't about an AI going rogue. It was about an AI agent getting framed by weak instructions. It had no clarity, no boundaries, and no lifeline. It was hired to babysit, and the "make sure the kids don’t play with the chainsaw" rule wasn't in the note.