PYX data product agentic engineering
← Field notes·07·Discipline·6 min

Vibe analytics is what your Genie pilot actually was

Ask a question, get a number, accept it because it looks about right. Software named this two years ago and built a discipline to escape it. We haven’t.

In February 2025 Andrej Karpathy described a way of programming where you give in to the vibes and forget the code exists. The term stuck because everyone recognised themselves in it. Analytics never got the equivalent, and we needed it more.

Here’s the loop. Someone types a question into a chat box. A number comes back. It looks about right, so it goes in the pack. Nobody reads the SQL underneath. Nobody can say what grain it aggregated at, or which accounts it quietly dropped. A month later someone asks the same question slightly differently and gets a different number, and now there are two, and no way to adjudicate.

That’s vibe analytics. And the reason it’s worse than vibe coding is straightforward: bad code crashes. A bad number presents.

The spectrum, not the binary

Google’s SDLC paper makes a point I’d borrow wholesale. Vibe coding and agentic engineering aren’t opposites, they’re ends of a spectrum, and the differentiator isn’t whether you use AI. It’s how much structure, verification and human judgement surrounds the output.

Slide that onto our world and the ends look like this. At one end, a question in a box and a number you eyeball. At the other, a requirement kept as written, grilled into typed claims, with grain and exclusions ruled on by a person, controls bound to those claims before anything runs, and evidence retained that you can re-run in eighteen months.

The right spot on that spectrum depends on the stakes. A curiosity on a Tuesday can be pure vibes. A number a regulator sees cannot.

Which is a more useful conversation than the one most teams are having. Nobody needs to ban the chat box. What they need is an explicit line: these questions, these environments, this discipline. The teams that keep the boundary blurry ship prototypes by accident. In data, shipping a prototype by accident looks like a figure in a board pack that nobody can reproduce.

Say it out loud in the room

Try this test. Tell your CFO the revenue exposure number on slide four was vibe analytics. Watch what happens to the room. Then tell them the same number came out of a stated grain, a stated exclusion set, a control that ran last night, and a verdict you can reproduce. Same number. Completely different conversation, and the difference is entirely in what surrounds it.

That’s not a technology gap. Every one of those pieces is buildable today. It’s a vocabulary gap. We’ve had no name for the disciplined end, so there’s been nothing to aim at and nothing to be caught short of.

Where to draw your line

Name the tiers, in writing. Exploration, internal, and regulated. Which questions live where, and which discipline each one earns. Vague boundaries are how a Tuesday curiosity ends up in a rate filing.
Set the bar at the check, not the demo. A demo proves it worked once. A bound control proves it holds. Ask for the control before you ask for the chart.
Make “I don’t know the grain” a legitimate answer. Right now the fastest path through most organisations is to guess and stay quiet. That’s a culture problem before it’s a tooling one.

So the next time someone demos a copilot answering a question in four seconds, don’t ask whether the number is right. Ask what would have had to happen for it to be wrong, and whether anything in the room would have caught it.

Frame borrowed from The New SDLC With Vibe Coding, by Osmani, Saboo & Kartakis, Google, May 2026. The mapping onto analytics is ours. See the full remapping.
Start a conversation

Have a data problem in mind?