Count the standards. AdCP, launched in October 2025 by a consortium including Scope3, PubMatic and Yahoo, gives buying and selling agents a shared language for discovering, comparing and activating media. ARTF arrived a month later, the IAB Tech Lab's answer, rebuilding real-time bidding so agents can operate inside the existing pipes. UCP, donated by LiveRamp to the Tech Lab and since rebranded, handles how agents pass audience signals to one another. The Trade Desk has its own kit. Underneath all of it sits MCP, the general-purpose protocol the wider AI industry already runs on. Eight months, five frameworks, one trade-body turf war, and an agentic era declared from every conference stage in the industry. Every one of these efforts answers the same question: how should two agents talk to each other? It is a real question and the engineering is real. It is also the easy half of the problem. Here is what a successful agentic negotiation produces under every one of those protocols: an agreement that each side writes down separately. The buyer's agent records the deal in the buyer's system. The seller's agent records it in the seller's. Two copies, two databases, two owners, no shared version of the truth. The protocols specify the conversation in exquisite detail and say almost nothing about the record. Which means the agentic stack is currently being built on the same bilateral bookkeeping that made the human stack unauditable. That bookkeeping has a track record. When ISBA and PwC attempted to follow advertiser money end to end through the supply chain in 2020, they could match buy-side records to sell-side records for 12 per cent of impressions. Twelve. Every participant kept its own books, the books did not agree, and 15p in every pound could not be accounted for at all. Nobody designed that outcome and nobody profited from it deliberately. It is simply what separate ledgers do at volume, and the study needed eighteen months of forensic accounting to establish even that much. Now run the same architecture at machine speed. Somewhere inside every holding company there is a team whose actual job, whatever their titles say, is making two systems' numbers agree: pulling logs, chasing discrepancies, adjusting pacing because the delivery count in one platform has drifted from the count in another. Alkimi simulated what happens to that work when agents take over the trading, and the disclosure comes first: the simulation is Alkimi's own, published in its May 2026 whitepaper, and vendor research deserves a discount, so apply one. Across 90,202 simulated transactions, agents keeping bilateral records generated 4,223 pacing interventions over a 90-day run, every single one triggered by drifted baselines rather than genuine delivery problems. The agents were not malfunctioning. They were responding correctly to records that had quietly stopped matching. Discount the precise figures as hard as you like. The direction survives, because the direction is just the ISBA finding restated: separate books drift, and everything downstream inherits the drift. So why has nearly all the standardisation effort gone into the conversation rather than the record? Because negotiation demos. A handshake between two agents fits in a keynote. A reconciliation queue does not. Discovery, comparison, activation, the phases AdCP formalises, are phases you can show an audience. The moment ninety days later when two systems disagree about what was delivered has no demo, no launch event and no press release. Investment follows the demo. It always has. The uncomfortable conclusion is that the protocol wars are a competition over the half of the problem that was always going to get solved. Agents were going to learn to talk regardless; the entire AI industry is building that layer for every sector simultaneously. What no general-purpose protocol will hand advertising is the other half: a record of what was agreed that both sides share, that neither side can quietly amend, and that a third party can audit without asking permission. That is not a messaging format. It is settlement infrastructure, and it is the only part of the agentic stack that a procurement department, an auditor or a regulator will ever ask to see. The winner of the protocol wars will not be whoever builds the best handshake. Handshakes are already commoditising. It will be whoever ships the boring half first.
3 September 2026