Programmatic advertising already had autonomy of action. A bidding engine set prices millions of times a second with no human approval step required, and has done so for well over a decade. The shift agentic advertising actually makes is different, and understanding the difference is the only way to evaluate a vendor's claim honestly. The shift is from autonomy of action to autonomy of agreement: two agents negotiating and closing a deal on terms neither side's humans confirmed in advance. That is genuinely new. It is also what creates a record requirement the old model never needed.
TL;DR. Programmatic automated the transaction inside human-set rules. Agentic advertising automates the agreement itself, which is why the category needs infrastructure programmatic never did. When two humans agree a deal, one record naturally emerges from the negotiation. When two agents agree, each writes to its own system, and those records can diverge from the moment the handshake completes. Published simulation research across 90,202 agent-to-agent transactions found that two agents which had just agreed the same deal recorded its terms differently in 95.3% of cases under separate record-keeping. The category is not programmatic running faster. It is a bilaterally agreed commitment between autonomous parties, and that changes what the infrastructure has to guarantee.
What was programmatic's autonomy, and why does it not count as agentic?
Programmatic's autonomy was autonomy of execution. A trading desk configured targeting, budget floors, bid ceilings, and frequency caps. Once those rules were in place, the platform bought autonomously: it submitted bids and won or lost impressions at a speed no human could match, without asking for sign-off on each decision.
That autonomy was always one-sided. The buying algorithm acted; the seller did not negotiate back. The auction cleared in milliseconds against fixed rules on both sides, and the only thing being automated was the matching of a bid to an impression. A human had decided in advance how much to pay, what to target, and what to exclude. The algorithm executed that decision at scale.
Autonomy of action of this kind has been the foundation of programmatic for fifteen years, and the industry built significant trust in it. The question agentic advertising is actually asking is different: not can software execute my instructions, but can software set the terms of a deal on my behalf.
What is autonomy of agreement, and why does it create a new problem?
Autonomy of agreement is what happens when a buyer's agent and a seller's agent negotiate a deal without their respective humans in the room. The buyer's agent carries a brief and a set of constraints. The seller's agent carries inventory and a pricing model. The two negotiate: volume, CPM, brand-safety conditions, frequency caps, delivery guarantees. They close on terms. The humans are notified of the result.
This is not execution inside human-set rules. It is negotiation on behalf of humans who were not present for the terms being agreed.
The record problem this creates is structural, not a technical flaw. When two people agree a deal, their agreement is the record. An insertion order is signed, emails confirm terms, and both parties start from the same facts. When two agents agree, each writes a record of the negotiation into its own system. Nothing in the current architecture forces those two records to match.
Research published with WPP modelled this directly across 90,202 simulated agent-to-agent transactions. Under separate record-keeping, where each agent maintained its own account of the deal, the two sides disagreed on at least one deal term in 95.3% of cases. Under a single shared record that both agents wrote to, divergence fell to 0.19%. The methodology is available from WPP Research.
Does the divergence compound, or is it one-time noise?
It compounds. In the same simulation, run to ninety days at holding-company scale, 677 million impressions settled with no agreed record of what had been delivered. The divergence does not announce itself: from inside each system, the account is internally consistent. The buyer's numbers add up. The seller's numbers add up. The two systems simply disagree about what the campaign was.
This is the failure mode programmatic never had to design against, because programmatic's transactions were discrete. An auction clears and the event is closed. There is no ongoing bilateral commitment to maintain. Agentic deals are different: they involve volume guarantees, pricing tiers, and terms that persist across a campaign flight. The longer the flight, the more opportunities for each side's record to diverge from the other's.
Shailley Singh of IAB Tech Lab, writing in August 2026, named the specific mechanism: "differences between LLMs degrading interpretation of the transactional context." Two agents can reach a genuine agreement in the conversational sense and still write down different structured representations of what was agreed, because the same words can be parsed into different meanings by different models.
What does this require of the infrastructure?
It requires what programmatic never built: a single record of the deal that both agents write to and both agents can query. Not two records reconciled at the end of a flight, but one record maintained in real time that is authoritative for both parties.
Brian O'Kelley of the Ad Context Protocol, asked whether a shared machine-readable record is a precondition for agent-to-agent buying to function as a market, answered in one word: "Yes." Singh, answering independently for the same ADOTAT investigation published in August 2026, described the requirement as "structured transaction objects, explicit transaction states, and systems of record that minimise hallucination or misinterpretation of context and that both sides can reference and audit." Both arrive at the same requirement from opposite sides of a competitive standards debate.
What should a buyer ask to distinguish agentic from automated-programmatic?
The question that cuts through most vendor presentations is simple: do the buyer's agent and the seller's agent reference one shared record of the deal, or does each keep its own?
A platform that routes bids through an agent interface but settles transactions through a unilateral record is automating execution, not enabling bilateral agreement. That is not agentic advertising in the sense that either the IAB Tech Lab or AdCP is building toward. It is programmatic with a conversational layer on top.
The capability to negotiate is what the demos show. The capability to hold both sides to what was negotiated, in a record neither side controls unilaterally, is what makes that negotiation a real transaction rather than a claim about one.
________________
This article references published simulation research into agentic deal reconciliation conducted by Alkimi Exchange and available from WPP Research, the IAB Tech Lab's published agentic advertising specifications, and statements by Shailley Singh of IAB Tech Lab and Brian O'Kelley of AdCP published in ADOTAT in August 2026. The simulation models the current specification architecture and is not an assessment of any specific production platform.