The Agent Sits With the Controller

Why I believe new agents will replace DSPs and SSPs — and who will force them to do it

Friday lunchtime, phone in hand, scrolling through the LinkedIn feed. In the second issue of his „Digital 6th Sense,“ Jens Depenau asks what programmatic will look like in 2030. Four scenarios are on offer: DSPs and SSPs build agents into their existing systems. New agents replace DSPs and SSPs. A hybrid model in which agents handle planning while platforms continue to handle delivery. Or the status quo prevails and agentic buying remains a niche.

I vote for the second scenario. The result that appears afterwards is telling. Integration and coexistence are tied at the top, with outright replacement trailing well behind. The status quo sits at zero percent. Nobody believes everything will stay as it is. But the majority apparently believes that the existing platforms will simply absorb the change.

That is a comfortable assumption. It is also the only answer that allows every party involved to keep its current business model.

What is technically happening right now

The standards Jens mentions in his post are real, and they are developing faster than this industry usually manages. In October 2025, a consortium including Scope3, Yahoo, PubMatic, Optable, Swivel and Triton Digital unveiled the Ad Context Protocol, or AdCP. It builds on the Model Context Protocol and the Agent-to-Agent protocol, and describes how buyer agents and seller agents find each other, negotiate terms and complete bookings. Publishers register the agents authorised to sell their inventory in a dedicated file, much like ads.txt.

The IAB Tech Lab consolidated its own work in February 2026 under the umbrella of AAMP — the Agentic Advertising Management Protocols. Version 2.3 followed in the summer, bringing a buyer agent, price guardrails and integration with major cloud platforms. And this is where I pause. Because the Tech Lab explicitly states that it does not want to replace OpenRTB, AdCOM and OpenDirect, but to „agentify“ them. The IAB’s standardisation path is therefore pretty much exactly scenario one: the old infrastructure gets an agent layer placed on top of it.

Anyone looking only at the specifications would have to vote for integration. I still think that is the wrong lens. The question of who buys media in 2030 will not be settled in the protocols — it will be settled in the finance department.

The problem is not technology. It is trust.

Anyone who has worked in the programmatic business long enough knows the feeling of holding a campaign invoice and not being able to say with any precision what was actually bought. The supply chain between advertiser and publisher has grown over years, layer by layer: agency trading desk, DSP, data provider, verification, exchange, SSP. Each of those steps takes its fee and operates by its own logic. The major study conducted by ISBA and PwC in the United Kingdom in 2020 found a gap that could not be attributed to any single actor. Roughly one seventh of the budget was simply unaccounted for.

I have used the analogy before, but it still holds: programmatic advertising functions in part like the high-frequency trading Michael Lewis described in Flash Boys. The opacity there is not an operational accident. It is part of the business model. Anyone sitting between buyer and seller — the only party who can see both sides — profits from precisely that information advantage.

For a long time, clients accepted this because there was no alternative. That is changing. Not because marketing departments have suddenly decided they want to understand adtech, but because finance departments are looking more closely. Media budgets are, in many companies, one of the largest freely controllable cost items. A CFO who insists on scrutinising every supplier in procurement will not indefinitely accept a line in the digital media budget labelled „Technology Fees“ that passes through without comment.

The agent on the right side of the table

This, for me, is the crux of it. An agent is not interesting because it books faster than a human. It is interesting because it can audit continuously. An agent on the client side can handle planning, purchasing, ongoing campaign management and, at the end, the financial reconciliation — as a single unbroken chain, not as a sequence of handoffs between departments and service providers. It monitors every campaign, reconciles every delivery against the booking and can say at any moment what share of the budget ended up with which publisher.

This only works if the agent belongs to the client. An agent integrated into a DSP works for the DSP. It will respond more helpfully than the old interface, it will write reports in full sentences. But it remains part of the same system whose margins it will never question. Scenario one ultimately means that the same black box can now talk.

If, on the other hand, a client’s buyer agent negotiates directly with a publisher’s seller agent via an open protocol, the question becomes unavoidable: what are the intermediary layers still needed for? Of course an ad still has to be technically served — somewhere adservers are running, somewhere milliseconds are being counted. But delivery is infrastructure. The DSP as the place where purchasing decisions are made and margins are extracted is something else entirely. And it is precisely that role which I believe can be replaced.

Data decides, not the model

One point receives far too little attention in the debate about agentic buying: the question of what these agents actually work with. An agent is only as good as its knowledge of the world in which it operates. And the knowledge floating around in the programmatic world is contaminated.

Historical campaign data contains years of ad fraud. Bot traffic counted as reach. Made-for-advertising sites whose click-through rates looked so promising that optimisation algorithms kept pushing more budget their way. Inventory sold through circuitous routes until nobody could trace its origin. Feed an agent this legacy data and you are essentially teaching it to procure fraud with maximum efficiency. An agent that learns from poisoned data is not a watchdog — it is a very diligent accomplice.

That is why I believe in agents that operate on the client’s own RAG foundation. A knowledge base built from verified results, the client’s own contracts, audited publishers and business effects that have actually been measured. Without the fraud-tainted legacy campaigns that have been accumulating in vendors‘ systems for years. It is more laborious than simply adopting a ready-made platform model. But it is the only path on which automation leads to greater control rather than faster loss of it.

From Report to Knowledge Graph

A knowledge base alone does not create transparency, however. It makes the agent smarter, but not the human who is supposed to look over its shoulder. That requires a second layer — and this, to my mind, is the genuinely new element: the knowledge graph.

Classic campaign reporting is a table. Rows of placements, columns of impressions, clicks and costs — sortable, exportable and virtually unreadable. What the table does not show are relationships. It does not reveal that three apparently independent news sites belong to the same operator. It does not reveal that part of the inventory passed through two resellers before it ever reached the SSP. And it certainly does not reveal that a platform used for buying belongs to the same holding company as the agency that recommended the buy.

A knowledge graph does not think in rows — it thinks in nodes and edges. Every advertiser, every agency, every holding company, every DSP, every SSP, every publisher, every individual domain is a node. The edges describe what happens between them: budget flows from here to there, this domain belongs to that publisher, this seller is authorised for that site, this platform is a subsidiary of that holding company. The underlying data is largely sitting in the open — it has simply never been brought together. It lives in publishers‘ ads.txt files, in SSPs‘ sellers.json files, in the supply-chain declarations that every bid request can carry, and in future — under AdCP — in publishers‘ agent directories as well. Add to that commercial registers and ownership structures. And the client’s own booking data, which records exactly how much money actually went where.

Lay all of that on top of one another and a picture emerges that has never existed before. You no longer see budget as a lump sum but as a current that branches, narrows at junctions and arrives at places you would never have expected. You see clusters of sites that all belong to the same operator and all have suspiciously high click rates. You see how many intermediary stops separated your euro from the publisher, and at which of those stops something was retained along the way. And you see conflicts of interest that appear in no agency contract but are plain to see in the ownership structure.

On my own blog I have been working for some time with a knowledge graph that maps how texts, topics and people connect with one another. The effect is always the same: relationships that you would never notice in a list jump out at you the moment you visualise them. Applied to a media budget, that means a controller no longer needs to understand how OpenRTB works in order to follow where the money went. She simply has to look.

For the client-side agent, the graph is more than an attractive interface. It is simultaneously its memory and its verification framework. Before it books, it can query the ownership structure behind an offer. While the campaign is running, it can detect when budget is migrating towards a cluster that has previously been flagged for fraud. And afterwards it can not only report what happened but show the route by which it happened. That is precisely where the difference lies between such an agent and one that is built into a DSP. The latter would have little interest in a graph like this — because above all it would make one thing visible: the DSP itself.

Where My Argument Gets Shaky

I do not want to pretend this is a settled matter. Established contracts are stubborn, framework agreements between holding companies and platforms run for years, and the real-time infrastructure does not vanish overnight. The IAB is deliberately building on existing foundations, and the major platforms have every interest in agents docking with them rather than routing around them.

The greater risk, however, lies elsewhere. Agents can themselves become a black box. If clients do not control their own agent but purchase it as a service from a new provider, nothing has been gained. In that case a new intermediary simply occupies the old seat, and scenario two is nothing more than scenario one with a new logo. The knowledge graph, too, is only as honest as its sources. Ownership structures are often deeply nested, sellers.json entries are incomplete, and whoever maintains a graph also decides what becomes visible within it and what does not. The transition I believe in presupposes that protocols remain open, that every decision the agent makes is logged and verifiable, and that the data sits with the client.

That is a demanding condition. But I believe the pressure to meet it will not come from adtech this time, where change tends to be packaged as a feature update. It will come from the finance department. And that, as a rule, has the longer staying power.

Somewhere right now, a controller is sitting in front of a spreadsheet with a line item called „Technology Fees.“ She has been waving it through for years. In 2030, she clicks on that line, and the spreadsheet unfolds into a web of holdings, platforms and domains, through which her budget runs as an ever-thinning line all the way to the last publisher. At one point, the line branches off to where the agency itself is sitting.

This English version was created with the help of AI (translated from the German original) and may differ slightly from a professional human translation.