The agentic A/B test doesn't exist.

So at Sharmin's dawat last Friday, the one in Road 17 with the marble lobby her husband polishes as if Queen Elizabeth is coming, my son's friend Tanveer cornered me near the kacchi station. He works in some Gulshan firm that does "agentic commerce." He explained, with great patience, that very soon I will not have to do my own shopping because an "agent" will negotiate with the seller's "agent" on my behalf. "Baba, you are describing the kameez man at New Market. Why are you so excited?"

He blinked. Sharmin's mother in law looked up because she smelled a story coming.

The kameez man at New Market teaches more than he sells. You point at the green silk with the silver thread on the yoke and ask the price. He says four thousand two hundred, you say one thousand eight hundred, everyone laughs including the boy fanning the flies. He pretends to call his cousin in Narayanganj who set the wholesale rate. None of those numbers are real. The price gets discovered through ninety seconds of theatre, with an interlude where he insists the fabric is "imported," which we both know means Mirpur. None of it is random, though I doubt Tanveer's people would recognise it as anything close to a "control group."

Tanveer kept saying "A/B test" as if it were an incantation. "Auntie, you show variant A to half the visitors and variant B to the other half, whichever wins, you ship." I said, "Baba, this is what your grandfather Ronald Fisher figured out in his 1935 book, na?" The whole thing only works if the assignment is actually random and every visitor is a fresh draw, like pulling rasgolla from a tin without looking. Tanveer looked surprised that I knew the man's name. I told him my son sends me articles. He does not, but Tanveer didn't need to know that.

Here is where Tanveer started losing the thread and I started taking over while he ate his kacchi. In this new world the buyer is a piece of software which has read the seller's "robots dot text" and six months of price history. Its owner has told it "get the detergent under five hundred forty taka per kilo." The seller also has software, defending margin within bounds set by some VP in a meeting nobody remembers. Variant A shows 540, variant B shows 510 with a "limited offer" badge. With human shoppers this is a clean little experiment. With software, the buyer reads the badge as a token, classifies it, discounts the signal because it has done this on ten thousand websites this week, and moves on. Worse, the buyer remembers your price was 530 last Tuesday, so your "random assignment" is being observed by something that keeps a diary. You haven't randomised anything, you've labelled two boxes the buyer can read. "Auntie, you are making it sound like a problem," Tanveer said. "Baba, I have not even started."

Once they figure each other out, the dashboard goes flat, and the variants are still doing something the whole time. It's the system that has settled into what mathematicians call a Nash equilibrium, a posh way of saying both sides have stopped trying because trying makes them worse off. The line oscillates for a few days, then settles into something that looks calm and is actually exhausted. Sharmin's mother in law, who has watched her grandson lie to his mother by looking healthy for thirty years, nodded at this.

So what should the marketing team do, I asked. Tanveer said his team is "iterating on test design." I told him to stop iterating and start thinking like a shopkeeper. There is a whole literature on bandit problems with many arms and adversarial feedback, two decades old, which marketing people ignored because human shoppers were not adversarial in any interesting way. Buyer agents are, not out of malice but because their owner's goal function was never yours to begin with. The smart teams are also borrowing from robotics, building a population of synthetic buyers with different prompts and budgets, letting them shop in a simulator, and watching what the seller does. Cheaper than a real test, and the CFO will notice it on the line item where the AWS bill used to be.

The other move, which I made Tanveer write on a napkin, is to stop asking "did this user convert" and start asking what the margin distribution looks like across ten thousand simulated buyers, and where the fat tails are. What comes back is a shape, a curve with fat tails on one end, and that is a much harder thing to explain to a VP who has been promoted on single numbers for fifteen years. The harder sell, I told him, is explaining next quarter why conversion has been stable for six weeks while the buying agents quietly learn his entire playbook.

I keep coming back to my kameez man, because everything new in this city eventually turns out to be something old wearing different clothes, and what he runs is a policy decades older than any optimisation framework in some Singapore handbook. He smiles at the regular and holds the line on the tourist whose grip on the fabric tells him she has no idea what silk weighs. He drops a hundred taka if the buyer turns to leave, because thirty years of these afternoons have taught him half the buyers come back inside. Every bit of it runs on full memory of the last visit and full intention of being there for the next, which is closer to negotiating than anything you could call randomised.

That, I told Tanveer as Sharmin brought out the firni, is where his industry is going whether his Monday planning meeting knows it or not. The experiment has not died, it has stopped being something you set up on a Monday and read on a Friday, and become something you negotiate every time the agent shows up. Tanveer asked if he could quote me, and I told him only on the condition that he credited the kameez man at New Market. Anyone who has bargained at Banga Bazar during an 8pm loadshedding, watching the ceiling fan slow and the shopkeeper pretend not to care, already knows a conversation is not a controlled trial.