A Dutch agency has published the architecture behind its agentic buying stack, and the diagram makes an argument the rest of the market has mostly avoided: the problem with agentic media buying is not the agents. It is the chain of command they inherited.

Lars Postmus, owner of Draft Digital and Draft Media, set out the case in a LinkedIn post carrying a relative timestamp of two weeks, placing publication in mid-July 2026. The post was accompanied by a two-slide diagram published under the draftdigital.nl domain. One slide is labelled "The old model, reskinned". The other is labelled "The Draft cockpit". Between them sits a claim about cost, control and headcount that touches every agency currently running agentic pilots.

The trigger was mundane. According to Postmus, he and three colleagues at the agency - Floris Bruin, Barbara Palfenier and Jasmijn Kruis - reached their Claude usage limits several times in a single working day. "Not because something broke," Postmus wrote. "Because we were building too hard. Good problem. But it points at a real one."

That real one, in his framing, is not a capacity issue. It is a structural one.

The chain that agentic tooling has so far preserved

The first slide maps the workflow that most holding companies and independents still run. A strategist briefs a comms planner. The comms planner briefs the specialists. Each specialist then buys on the platform they own.

The diagram names five specialist roles: a social buyer, a search buyer, a programmatic buyer, an audio and DOOH buyer, and an offline buyer. Beneath them sit twelve named endpoints. Social runs on Meta, TikTok and Reddit. Search runs on Google. Programmatic runs on PubMatic, VDS, Mediahuis and DPG. Audio and DOOH run on Triton and BroadSign. Offline runs on Ster and Ad Alliance.

The annotation on that slide is the point of the whole exercise: every box is one MCP call, every single time.

The Model Context Protocol is the connective tissue underneath most current agentic advertising work. It is Anthropic's open standard for wiring language models to external systems, and it is the foundation on which the Ad Context Protocol was built when AdCP launched on October 15, 2025 with six founding members and 23 additional launch participants. Platform after platform has since exposed an MCP surface. Amazon opened a closed beta in November 2025. Meta, Google and a long list of measurement vendors followed through 2026.

Each of those connections is useful in isolation. Stacked into a legacy workflow, they multiply. Twelve endpoints queried by five separate buying functions produce twelve separate context loads, twelve separate authentication paths and twelve separate rounds of instruction. The bill arrives in tokens.

"Many buyers. Many tools. Tokens burn on every call, and the strategy quietly fades on the way down. New tech, same old problem," Postmus wrote.

Two distinct complaints sit in that sentence. One is a cost line. The other is what happens when a plan is translated four times before anyone spends money against it.

What the cockpit rearranges

The second slide inverts the hierarchy rather than compressing it.

The strategist sits at the top and, in the diagram's phrasing, stays in the lead for the plan. Four agents operate around that position: a Brand Agent, a Performance Agent, a Creative Agent and a Data Agent. Those four steer a single Buyer Agent. The Buyer Agent, in turn, talks to one Agentic Ad Platform described as one layer covering all channels. Below that layer, the five channel groups from the first slide reappear as feeds rather than as separate buying desks, each paired with an Optimisation Agent.

The arithmetic is the argument. One agent negotiating with one platform layer replaces twelve direct integrations queried independently. The slide records three claimed outcomes: the strategist keeps control, calls fall and tokens stay low, and the strategy stays in the plan.

None of those claims arrives with a benchmark. The material published contains no token counts, no cost comparison against the fan-out model, no campaign results and no named client for the cockpit itself. What it contains is a topology.

Where the specialists land

This is where the diagram addresses the question the industry keeps asking sideways.

"The specialists do not disappear. They move up and feed an Optimisation Agent per channel. Humans do the thinking, agents do the lifting," Postmus wrote.

The formulation is precise about what changes. In the old chain, a search buyer executed inside Google Ads. In the cockpit, that person supplies judgment to an optimisation layer that executes. The task moves. The expertise is retained as an input rather than as an operator function.

Whether that distinction survives contact with a procurement conversation is a separate matter. An agency that bills for execution hours and then automates execution has to find a different basis for the fee. Ari Paparo made a version of this point in July 2025, when he argued that agentic AI threatens traditional DSP business models and that the value proposition for media buying services would shift toward strategic planning and creative development rather than tactical execution. The cockpit diagram is, in effect, an operating model built on that premise.

What the token argument actually costs

The claim on the second slide, that calls fall and tokens stay low, separates into two parts, and the post moves between them. The incident that prompted it is a development constraint. Four people reached their Claude usage limits while building the cockpit. That is a subscription cap consumed by engineering work, and it carries no information about what a finished campaign costs to run. The slide's claim concerns production. The anecdote supporting it comes from the other side of the ledger.

Production figures are thin. At IAB Europe's live agentic showcase on April 14, 2026, one demonstration completed a full task across three agents working asynchronously in roughly 25 minutes, at approximately 50 euros in token and credit charges. Against a media plan of any size that is immaterial. Against the volume of briefs an agency processes in a quarter it is a line item that did not exist two years ago. Gartner has forecast that more than 40% of agentic AI projects could be abandoned by 2027, with escalating costs among the reasons.

The technical objection concerns what drives the bill. Call count is a weak proxy. Consumption tracks context volume, and one agent loading twelve platform catalogues into a single working context is not obviously cheaper than twelve agents each loading one. What consolidation removes is duplicated context: the brief, the client history, the brand rules, restated on every branch. That is a real saving, and narrower than the diagram implies.

Format matters at least as much as architecture. Cloudflare's Markdown for Agents, introduced on February 12, 2026, cut the token cost of serving web content to AI systems by roughly 80% by converting HTML to markdown on request. The company's illustration is blunt: a heading consuming three tokens in markdown burns twelve to fifteen in HTML before wrapper elements, navigation and scripts are counted, and its own announcement post measured 16,180 tokens as HTML against 3,150 as markdown. Twelve advertising endpoints of varying technical maturity would produce comparable variance inside any agentic stack, whatever sits above them.

Then there is the direction of travel, which cuts both ways. Alphabet reported that Gemini serving costs fell 78% during 2025. Cheaper inference does not reliably produce smaller bills. A Google engineer told a marketing audience in May 2026 that token consumption follows Jevons paradox, the nineteenth-century observation that efficiency gains raise total consumption rather than lowering it, and that tokens would end up throughout the workflow, attaching a visible cost to work that had previously been invisible. Cloudflare answered the same pattern by opening a public beta of dollar-denominated spend limits in its AI Gateway on June 5, 2026.

An architecture justified mainly on token economics rests on a number that keeps falling and a volume that keeps rising. The second half of the Postmus argument is sturdier. Strategy decay does not track the price of inference.

The standards backdrop

Postmus closed the post by naming where he expects the market to settle. "That is where AdCP and Agentic Advertising are heading. We are building for it now."

Both references are specific. AdCP is the Ad Context Protocol, governed by the nonprofit AgenticAdvertising.org, which is also listed among the affiliations on the Postmus LinkedIn profile alongside the Amsterdam University of Applied Sciences. The protocol defines nine core tasks covering the advertising lifecycle, from inventory discovery through campaign activation and reporting, and it has been contested since the week it appeared. AdCP divided the advertising industry over whether standardising the transaction layer addressed or deferred the structural problems in programmaticsupply chains. Ari Paparo published a detailed analysis on November 3, 2025, supporting parts of the specification while questioning the media buying applications. Transparency objections surfaced in parallel, with debate over what agent-mediated buying does to supply chain visibility.

The governance landscape has since fragmented and partially reconverged. LiveRamp donated the User Context Protocol to IAB Tech Lab in November 2025, positioning it as a data plane beside AdCP's control plane. IAB Tech Lab named its own umbrella initiative AAMP on February 26, 2026, having earlier opened an agentic RTB framework for container-based advertising. Prebid.org took stewardship of an open-source sales agent in January 2026 and placed publishers at the centre of its agentic roadmap. Scope3 co-founder Brian O'Kelley, who reposted the Draft Digital diagram, published adagents.json as a proposed replacement for ads.txt on March 30, 2026.

Affinity joined AdCP as a founding member in June 2026. What the protocol has not gathered is the demand side that matters most: Google, Amazon and The Trade Desk have not adopted it. That absence is material here. A single Buyer Agent talking to one agentic ad platform assumes that platform reaches the inventory the plan requires. Where the largest buying systems sit outside the protocol, the fan-out the diagram removes reappears elsewhere in the stack.

Whether the market data supports the premise

Two separate readings are available, and they do not agree.

The efficiency case has vendor evidence behind it. In a December 2025 campaign on PubMatic's AgenticOS, the agency Butler/Till reported 40% more impressions than comparable non-agentic activity, a 5.5x cost efficiency and a 98% average video completion rate, on a single test. PubMatic launched AgenticOS with live campaigns running on January 5, 2026. AdRoll and PubMatic connected agents to diagnose programmatic deal problems in April 2026. Draft Digital itself supplied the buy-side agent for what Broadsign described as the first fully agentic OOH campaign, announced on May 27, 2026, for Dutch charity lottery Lot of Happiness across Global Netherlands inventory, using AdCP as the communication standard between the two agents.

The counter-reading is more recent and less flattering. DataBeat's June 2026 US Programmatic Trends Report found that conventional programmatic buyers held a 13.4% CPM advantage over AI agents across the May 2026 data period, near parity rather than the multiples early vendor tests suggested. And the company that has pushed agentic advertising hardest has been shrinking: Scope3 cut engineering and sales staff in February 2026, five months after commercial team reductions in August 2025.

Adoption inside agencies remains uneven. Research covered in April 2026 found that 87% of agency professionals considered the traditional agency model broken, with agentic AI reaching 46% of surveyed agencies. Adoption clustered in low-consequence tasks: ideation at 86.9% and research at 84.0%, against media planning at 29.1% and media buying strategy at 22.1%. The stages where the cockpit diagram proposes to operate are precisely the stages where uptake has been slowest.

Why the architecture question matters now

A cluster of announcements on June 11, 2026 - Mediaocean's NIVO intelligence layer, Magnite's Orchestration coordination product and Teads' EngageOS - prompted the question of who controls the buying layer now. A week later, further launches from DoubleVerify and LiveRamp confirmed that agentic ad tech was moving on the buying layer directly, with Mediaocean survey data showing 60% of marketers planning to raise AI media spending in the second half of 2026.

Almost all of that construction has come from the sell side or from technology vendors. The Draft Digital slides come from a buyer, and answer a different question. Vendors are building coordination layers between agents. Postmus is describing where the human sits relative to those layers, and what becomes of the people who used to sit between the strategist and the spend.

Omnicom raised the same workforce implication from the holding company side, as noted when Big Tech recorded 150 billion dollars in first-quarter advertising revenue. The framing that automation frees teams for strategy is familiar from earlier waves of programmatic development. Whether the outcome preserves human roles at equivalent scale is answered by the cockpit diagram through assertion rather than evidence.

What the post does not establish

The material is a LinkedIn post and a two-slide diagram, not a product launch, a technical specification or an audited result. No token figures accompany the claim that calls fall and costs drop. No client is named for the cockpit. The Agentic Ad Platform at the centre of the second slide is not identified as a vendor product or as internal tooling. The five specialist roles are described as moving up without any statement about how many people that involves.

Engagement was modest at capture: 28 reactions, three comments and one repost, on a profile with 4,885 followers. The repost came from O'Kelley, whose company co-founded the protocol the post endorses. One public comment engaged with the substance. Nirar Punj, a performance marketing specialist, wrote: "This resonates so much with our own token challenges! Love the strategist-centric approach."

That is a small sample, and also a second practitioner reporting the same constraint.

The question underneath

Read against the AdCP debate, the diagram reframes rather than resolves the displacement question. The protocol was designed to let agents transact across platforms without bespoke integration work. If that succeeds, the operator function inside each platform interface loses its scarcity. What survives is judgment about allocation, positioning and creative direction, which the cockpit places at the top and calls the plan.

The counterargument is simple. Routing every channel through one buyer agent concentrates decision-making inside a system whose reasoning is difficult to audit, at a moment when the industry has not settled how agent-mediated buying reports its supply path. The transparency objections raised against AdCP in November 2025 apply with more force, not less, when the fan-out is replaced by a single layer.

Postmus framed the market's current activity generously. "Right now the whole market is testing Agentic," he wrote, before adding: "But most of these tests do not fix the real thing yet. The strategist problem."

The diagram is a claim about where control belongs. The evidence that it produces better outcomes, at lower cost, with the specialists intact, has yet to be published.

Timeline

Summary

Who - Lars Postmus, owner of Draft Digital and Draft Media in the Netherlands, publishing alongside colleagues Floris Bruin, Barbara Palfenier and Jasmijn Kruis. The post was reposted by Brian O'Kelley, co-founder and chief executive of Scope3 and a founding member of the Ad Context Protocol.

What - A two-slide architecture diagram and accompanying commentary arguing that current agentic advertising pilots preserve the briefing chain that causes both token cost and strategy decay. The proposed alternative places the strategist at the top, surrounds that position with Brand, Performance, Creative and Data agents, routes all activity through a single Buyer Agent into one agentic ad platform, and converts five specialist buying roles into inputs to per-channel Optimisation Agents.

When - Published on LinkedIn with a relative timestamp of two weeks at the point of capture, placing publication in mid-July 2026.

Where - The Netherlands, with the diagram published under the draftdigital.nl domain. The named endpoints span Dutch and international supply, including Meta, TikTok, Reddit, Google, PubMatic, VDS, Mediahuis, DPG, Triton, BroadSign, Ster and Ad Alliance.

Why - The material addresses the operating cost and control question that agencies running agentic pilots have not resolved. Twelve platform integrations queried by five separate buying functions multiply token consumption and dilute the original plan. The design also takes a position on displacement, holding that specialists move upward into optimisation roles rather than out, at a moment when adoption data shows agentic tooling reaching only 22.1% of media buying strategy work and independent measurement placing agent performance near parity with conventional buying.