"Agentic": the word your whole industry is shouting, and almost nobody can pin down

In adtech, the technology is almost never the problem. The operational layer always is. This series is about that gap. One buzzword at a time.

I counted once, in the run-up to one of the big industry weeks: eight adtech platforms shipped an "autonomous buying" or "agentic coordination" layer within days of each other. Eight. The verification players, the SSPs, the media-ops giants, all racing to plant the same flag before the rosé started flowing.

Here's what nobody on a main stage will tell you: no single one of those launches was the story. The fact that they all landed at once was the story, and it's not the story they thought it was.

Because when everyone ships the same thing in the same window with the same word stamped on it, that word has stopped meaning anything. So let's give it back a meaning. No jargon, no hype. Just what "agentic" actually is, when it's real, and when it's a slide.

What everyone's saying

"We're entering the agentic era." "Agents that plan, buy, negotiate, and optimize on their own." "If you're not agentic, you're behind." It's the loudest word in the industry. It has quietly become one of the most overused and least understood terms in the business. Both halves of that matter. It's everywhere, and almost nobody you'll meet can tell you where the autonomy actually starts and stops.

What it actually is

Strip the costume off and "agentic" means one specific thing: software that loops toward a goal and corrects itself as it goes, instead of running a fixed sequence and stopping.

That's the whole distinction. Normal automation runs a script: every Monday, pull this report, push this budget, in this order. Same input, same output, every time. It executes a sequence and stops. An agent is handed a goal, "keep this campaign pacing on target", and runs a loop: it acts, checks the result against the goal, and tries again, adjusting each time, until it gets there.

The simplest way to feel the difference: automation is a recipe, an agent is a cook. The recipe runs the same steps no matter what. The cook tastes, notices it's under-seasoned, and adjusts, working toward "this should taste good" rather than "follow steps one through five." Both have a human deciding what's for dinner.

And here's where most explanations get it wrong. It's not that the agent "decides for itself" and you don't. A human sets the goal either way: the objective, the limits, what it's allowed to touch. Put an agent on a tight leash, make it ask before every move, and it's still an agent. What makes it one was never independence from you. It's the loop.

And no, it isn't about speed. Automation is already fast. A faster sequence is still a sequence. The loop is the whole ballgame. Everything else is detail.

Why it matters when you're the one spending the budget

Two things are true at once, and you need to hold both or you'll get played.

First truth: the mechanism is real. Strip away the marketing and something genuine is happening: buyer agents and seller agents are now negotiating actual buys, not demos. Do the deals you read about exist? Be honest about what you're looking at: almost every "first agent-led deal" headline is a press release from a party with every reason to inflate it. The numbers attached (5x cheaper, 45% more impressions) are vendor-stated, unaudited, and built to sell you a product. Throw all of that out. What's left underneath, once you've cleared the marketing, is a hard core you can't dismiss: agents on both sides of a transaction, negotiating real inventory, with money on the line. The mechanism is real. It's the performance that's being oversold. Pretending the whole thing is vapor is just denial running the other direction.

Second truth: the people who spend the money aren't buying the hype, and they're right not to. The pattern shows up in survey after survey. Record shares of marketers planning to increase AI spend, while the share who believe AI actually transforms their workflow stays low, sometimes falling even as budgets climb. More money, lower expectations. The industry analysts who track hype cycles have put agentic AI right at the peak of inflated expectations. And the blunt verdict from people walking the trade-show floors keeps coming back the same: buyers find agentic AI more interesting than urgent.

So where does that leave you? With a job most people are skipping: telling the real thing from the costume, on a case-by-case basis, before you wire budget or signatures to it.

What nobody selling you "agentic" will admit

Most of what's being sold to you as agentic isn't.

I'll be blunt, because that's the point of this series: a large share of the "agentic" products being paraded around are ordinary automation with a new label glued on. A scripted workflow (fast, useful, completely pre-defined) rebranded as an autonomous agent because the word moves budgets. It's the same move as last decade's "AI-powered," and the decade before's "programmatic everything." The label changes; the trick doesn't.

Here's the test that cuts through it, and you can run it in any vendor meeting without being technical:

Ask what makes it stop. Automation stops when it runs out of steps. An agent stops when the goal is actually met, and keeps working until it gets there.

If it notices it fell short and works the problem again, on its own, you may be looking at something genuinely agentic. If they pivot to how fast it is, how much manual work it removes, how clean the dashboard looks, that's automation wearing a costume. Useful, maybe. But don't pay an agentic premium, or hand over agentic-level trust, for it.

And there's a real cost to getting this wrong, beyond the price tag. The sharpest worry in the market isn't that agents won't work. It's that "agentic" becomes the next must-have label every intermediary slaps on, rebuilding the same opaque middle layer that has always quietly skimmed a buy, now with per-transaction AI fees stacked on top. Faster commoditization, dressed as progress.

The real question nobody's asking

So here's what actually keeps me up, and it's the whole reason I'm writing these.

Those eight platforms and all the ones that follow them are selling you the same thing: the agent, the smart part that does the work. But an agent is only as good as what it's working over. Point a brilliant autonomous system at a noisy, incomplete, half-mapped supply chain and you don't get smarter advertising. You get your blind spots automated, made permanent, and more expensive, because now they run at machine speed with nobody watching.

That's the pattern, and it's the thing to carry into every acronym I break down next: in adtech, the technology is almost never the problem. The operational layer underneath it always is. The agent is the part everyone's selling. What it stands on is the part that actually determines whether it helps you, and it's the part nobody's putting on a slide.

Next in this series, I take the first brick out of the "agentic" wall and show you what's really under it: MCP, the protocol everyone name-drops and nobody can explain.

---

Maxime Khalfallaoui, Supply Finder. I came up on the SSP side. Now I help agencies find their way through it.