By day 57 of the simulation, the agent had run twenty campaigns and was optimising with complete confidence. It was also wrong about all of them. Not by much on any single deal, but every figure it had learned from had drifted from the start, so the more experience it gained, the surer it became of numbers that were never real.
The agent was not the problem. What it ran on was. That is the part of going agentic almost nobody is selling you, and the part you will still be living with in three years. Which is why the question that matters in 2026 was never which agent to buy.
Going agentic is not a shopping trip
The whole industry decided to go agentic at once. IAB Tech Lab spent the back half of 2025 and the start of 2026 turning its standards into agent standards, unveiling an agentic roadmap in January and naming the umbrella AAMP in February. A rival camp shipped the Ad Context Protocol in October 2025, six founding companies, built on Anthropic's Model Context Protocol. Even the man setting much of the agenda, IAB Tech Lab's Anthony Katsur, has called agentic a "shiny penny".
The vendors moved faster than the committees. PubMatic ran what it billed as the industry's first fully autonomous CTV campaign in December. Google put an Ads Advisor inside Google Ads. ChatGPT started serving ads in February at a reported $60 CPM. At CES the pitch was everywhere, and everywhere it sounded inevitable. One 2026 field guide summed up how buyers actually felt about it: "more interesting than urgent".
Here is the part the demos skip. Going agentic is not a purchase. It is a change to how you operate. The thing on the shelf is somebody else's operating model with a login on the front. An agent that came out of a box is not an edge, because the shop across the road bought the same box, trained on the same standard tasks, reasoning from the same defaults. You did not buy an advantage. You bought parity, and you paid a margin for it.
Which is why the interesting question splits in two.
The question that split in two
For fifteen years the question was: which platform do I license? Going agentic breaks it into two questions that used to be one.
The first is what gets built. Not which template you switch on, but what an agent shaped to your brief actually does. Your constraints. Your definition of a good outcome. Your rules about where it will and will not spend, and what it does when it hits the edge of them.
The second is what it runs on. The ground beneath the agent. Where the deal gets recorded, who can read that record, and, when the agent gets something wrong at three in the morning across four hundred deals, whether you can prove what it was told and put the fault where it belongs, or whether you are the one left holding it.
Most of the market answers the first question with a demo and the second one not at all. It sells you the agent and quietly rents you the ground it stands on. You get a clever buyer and a locked box underneath it, and the two arrive bundled so tightly you never think to pull them apart.
Pull them apart.
The ground you don't own
Say the agent is good. Say it genuinely buys well. It still has to run somewhere, and where it runs is usually infrastructure you do not control, keeping a record you cannot independently read. That sounds like a technicality until you watch what happens to the record over time.
Alkimi Research modelled it, with WPP, across 90,202 simulated deals. When two agents negotiate and each keeps its own record of what was agreed, the records drift apart. Not through error. Through the ordinary mechanics of two systems each maintaining its own version. In the simulation, 95.3% of transactions disagreed on at least one deal dimension under separate record-keeping, against 0.19% when both agents wrote to a single shared record. Over a simulated 90-day run at agency holding company scale, 677 million impressions had no agreed delivery record at settlement. The buyer paid against its numbers. The seller invoiced against theirs. And 76.7% of deal value came out structurally mispriced while still passing standard reconciliation checks. The drift compounds, which is what the agent at the top of this email was living. Its confident, wrong, day-57 self had made 4,223 pacing corrections across that study, against zero under a shared record, every one of them a fix for a problem that existed only because the two sides never kept the same numbers. All of that is simulation, not a forecast about any live platform, and the paper is careful to say so.
None of it is new. It is the reconciliation gap the industry has lived with for two decades, the one the ISBA and PwC supply chain studies put a number on: at the premium end, roughly half of advertiser spend reaching the publisher, six to ten intermediaries stood between the two ends, and nobody able to reconcile the books without everybody's cooperation. That gap was survivable when humans argued it out slowly after the flight. An agent removes the "slowly". There is no after-the-flight when the next thousand deals have already cleared.
The industry has quietly agreed the ground doesn't matter as long as the agent demos well. That was always a bet. Agents are the thing that calls it.
So
So the choice was never which agent is cleverest. It is what gets built for you, and whose ground it stands on. Two questions, two decisions, both yours to make if anyone lets you make them.
And here is the objection you should be forming already. Fine, but if the answer is some startup's settlement infrastructure, have I not just swapped one black box for another, from a smaller company, with my ad ledger sitting somewhere I would struggle to defend to my board? Hold that thought. It is the right one. I work at that company, so read the next bit with the objection loaded.
Bring the brief
Alkimi does both halves. Bring the brief and we build the agent to it, or configure one to it, depending on what the brief needs. Some briefs want bespoke logic. Most want a capable agent pointed hard at one outcome, with constraints that hold under pressure. Either way it is shaped to you, not lifted off a shelf and renamed.
Now the objection. The record does not sit in Alkimi's private database for you to request screenshots of. It sits on shared settlement infrastructure both sides write to and neither can quietly edit, the same unglamorous idea the equities markets reached for when they stopped keeping two separate sets of books. The chain underneath it is the mechanism, not the pitch. And the real test of whether a record is yours is whether you can walk away and still read it.
Any performance claim I would make for it comes with its method attached or it does not get made: an uplift number carries its third-party design in line, a geo-holdout through Measured, an A/B test, a geo-lift. Anything from the simulation stays labelled simulation.
If you would rather see it against your own supply chain than take my word for it, that is a conversation, not a signature. Lauren runs those: lauren@alkimi.org.
The one that stays
The demo will keep getting better. The close will keep getting faster. None of that answers the only question you have to live with once the money has moved and something has gone wrong.
So carry it in with you. Not how clever the agent is. When it makes the wrong call on your budget, whose record decides whose fault it was? Everything else is a feature. That one is the whole job.
Interesting reads
IAB Tech Lab announces an agentic roadmap, but don't expect any love for AdCP (The Current). A clean read on the politics underneath the standards, and the pointed silence around who launched first. https://www.thecurrent.com/data-privacy-iab-tech-lab-ai-agentic-roadmap-adcp
Agentic Ad Tech Confusion: Revolution or Evolution? (ExchangeWire). The clearest framing I have seen of the build-from-scratch versus build-on-what-works split now dividing the industry. https://www.exchangewire.com/blog/2026/01/14/agentic-ad-tech-confusion-revolution-or-evolution/
Agentic AI in Advertising: A 2026 Field Guide (TensorOps). The sober version. What agents genuinely run end to end today, and what still needs a human who understands the business. https://tensorops.ai/labs/research/agentic-ai-advertising-2026-field-guide
Related Reads: