A media buyer running a Performance Max campaign in 2026 has already handed a lot of the job to software. Google's machine learning picks the placements, sets the bids, assembles the creative, and decides which audience signal to chase. The buyer sets a goal and a budget, uploads assets, and mostly watches. So when the industry started talking about agentic media buying, a fair first reaction was: did we not already do this? Is an agent just a better Performance Max?
It is not, and the difference is worth getting exactly right, because it changes who is running the campaign. Performance Max optimizes a campaign someone else built. An agent builds the campaign. That sounds small. It is the entire shift.
Origin: from automated to autonomous
Programmatic advertising has been automated since its birth. The point of real-time bidding and demand-side platforms was always to take the manual labor out of buying media, and by the early 2020s that automation had climbed steadily up the workflow. Bid management, audience modeling, budget pacing, and creative testing all moved from spreadsheets to machine learning. Google's Performance Max, launched in 2021, and Meta's Advantage+ pushed it furthest: one campaign type where the system decides where the ad runs across an entire property network.
But every one of those systems shares a hidden limit. They optimize inside a box a human drew. A person still defines the campaign structure, picks the objective, writes the brief, sets the constraints, judges whether the results were any good, and decides what to do next. The automation handles execution within those walls. It does not move the walls. eMarketer frames the current state plainly: generative AI is becoming the default for bidding, audience building, and creative variants across programmatic, but that is optimization, not autonomy.
Agentic media buying is the attempt to remove the box. An agent is software that takes a goal and works out the steps itself. Give it a campaign objective and a set of guardrails, and it plans the approach, queries for audiences, sources the inventory, decides what to buy, monitors the result, and adjusts. It does the planning and the judgment that automation always left to a person. This is the broader idea of AI agents, software that plans across multiple steps, holds context, calls tools, and acts toward an outcome without a human scripting each move, pointed at a media plan.
The honest line between the two: automation answers "how do I execute this campaign well." An agent answers "what campaign should I run." The first is a faster hand. The second is a different decision-maker.
Present: what an agent actually decides
The abstraction gets clearer with a concrete system. PubMatic launched AgenticOS on January 5, 2026, calling it an operating system for agentic advertising. Strip the branding and it is a set of specialized agents, each owning one part of a media buy, coordinated by a central buyer agent. The structure maps what an agent does that a Performance Max campaign does not.
There is an audience discovery agent that takes a plain-language description of who the campaign should reach and queries hundreds of data partners for matching segments, before any campaign parameters exist. An inventory marketplace agent takes the campaign intent, format, audience, and brand-safety needs, then assembles ranked inventory packages from the publisher network in real time. A media activation agent handles budget strategy, pacing, flighting, and deal targeting across thousands of publishers. An insights agent watches the live campaign and surfaces pacing anomalies before a human would notice them. Above all of these sits the buyer agent, which translates a high-level goal into instructions and orchestrates the rest.
Read that list against a traditional workflow and the shift is obvious. Finding the audience, sourcing the inventory, building the package, judging whether delivery is on track: those were the planner's and trader's jobs. The agent is not optimizing a campaign a human structured. It is doing the structuring. Scope3, one of the companies building the plumbing for this, describes the agent as an autonomous decision-maker that evaluates media opportunities against brand-specific goals rather than executing a fixed plan.
One technical point clears up a common confusion. A large language model cannot bid in a real-time auction. RTB resolves an impression in roughly 100 milliseconds, and no LLM responds that fast. So the agent does not sit inside the bidstream pricing impressions one by one. It works a layer up, setting the strategy, the parameters, and the inventory packages, while the existing high-speed bidder runs the actual auctions inside the agent's guardrails. The agent is the planner and the trader. The old machinery is still the auctioneer.
Present: the early results, and how much they prove
The first live campaigns produced numbers worth taking seriously and worth reading carefully. The clearest case study is independent agency Butler/Till, which ran an agentic campaign with PubMatic for brewer Geloso Beverage Group across connected TV, online, and mobile apps from December 2025 into January 2026. Digiday reported the results: buy-side costs, mostly the tech and DSP fees layered between buyer and publisher, fell 82 percent. Effective CPMs dropped 30 percent, and the campaign ran 40 percent more impressions than planned. Quality held up, with a 98 percent video completion rate and a made-for-advertising rate below 1 percent verified by Jounce Media.
PubMatic reports the pattern is holding as the program scales. It says it has run more than 30 fully autonomous end-to-end campaigns globally, and that every advertiser that ran one came back to run another. Its emerging revenues, the category that includes its newly launched AI products, grew more than 80 percent year over year in the first quarter of 2026.
Two cautions keep those figures honest. First, they come from the vendor or a partner the vendor enabled, on a small set of early campaigns: encouraging, but not independent, audited, or large-scale. Second, the 82 percent cut in buy-side cost is partly an artifact of the route. Buying through a supply-side platform's own agents removes intermediary layers a normal programmatic buy passes through. Some of that saving is the agent being smart. Some is the chain being shorter because the path is new. A buyer should not assume an 82 percent fee cut transfers to every agentic setup.
What the early results do prove is narrower and still significant. The slow, manual parts of a media buy, finding segments, sourcing and packaging inventory, setting up the campaign, spotting problems, can be compressed hard without obviously wrecking quality. Whether an agent should be trusted with strategic and creative decisions is a separate question, and the people running these tests are mostly answering it the same way.
Present: where the human line is being drawn
The Butler/Till campaign is the most useful detail in the whole story, because of how it actually ran. The agency's agent and PubMatic's agent talked to each other directly, passing targeting parameters and inventory options back and forth. But the human team wrote the client brief and approved the inventory selections before anything was bought. The agents handled the back-and-forth and the execution. People still owned the brief and the sign-off.
That is the line nearly every serious operator is drawing, and it holds across very different companies. Amazon's revamped Campaign Manager offers a Smart Mode where AI runs the campaign once goals are set, and an Expert Mode that keeps the human in control with AI as a support layer. The Trade Desk's Koa Agents are rolling out in phases that start with audience planning and inventory prioritization, the assistive end, before touching autonomous execution. The recurring position across the practitioner debate is that agents should accelerate the operational work while experienced people stay responsible for strategy, judgment, and accountability.
The skeptical case is sharper still. Christopher Francia of Attention Arc told Digiday that programmatic activation has too many variables to trust to LLM-based agents, and that an agent making a decimal-point error in a budget can be financially catastrophic. His point is not that agents are useless. It is that they are reliable on narrow, well-defined tasks and unreliable on open-ended ones, the pattern across agentic AI generally. The genuinely hard part is not the agent doing work, it is knowing which work is safe to give it. The answer that has emerged is a division of labor, not a handover: agents take the repetitive, measurable, reversible parts, and humans keep the brief, the creative, the strategy, and the final approval to spend.
Future and impact: standards, scale, and the honest risks
For agents to buy and sell media across companies, a buyer's agent and a seller's agent need a shared language, which is why a standards contest is running underneath all of this. The Ad Context Protocol, launched in October 2025 by Scope3, Yahoo, PubMatic, and others, builds advertising-specific modules on top of Anthropic's Model Context Protocol. The IAB Tech Lab named its competing initiative the Agentic Advertising Management Protocols, or AAMP, in February 2026, adapting its existing standards rather than adopting AdCP. Practitioners are openly unsure whether the two efforts are compatible or rivals. A split would slow everything down: agents that cannot speak the same protocol cannot trade. But the direction of travel is not in doubt even if the speed is. eMarketer's read is that full agentic programmatic is unlikely within 2026, but that more of the workflow becomes autonomous each quarter, starting with reporting and campaign operations. The question is no longer whether agents enter media buying. It is how far up the decision stack they climb, and how fast.
The risks deserve as much attention as the promise.
The first is a return of opacity. Programmatic spent a decade fighting to see where money goes, through supply-path optimization, curation, and transparency reporting. If a buying agent transacts with a selling agent through layers of automated negotiation, a marketer can lose sight of where ads ran, what data was used, and what fees were taken, faster than any human could audit. The pitch for agents is that they attack programmatic's oldest problem, the murky supply chain. The failure mode is that they rebuild it, quicker and harder to inspect. Which outcome a buyer gets depends on whether the agent will show its work.
The second is accountability. When an agent overspends, buys the wrong inventory, or runs an ad against unsafe content, who is responsible: the buyer who set the goal, the agency that deployed the agent, or the vendor that built it. Law firm Davis and Gilbert, reviewing the legal exposure of agentic advertising, is direct: regulators will treat an AI's decisions as the company's decisions, and the absence of human intent will not shield a brand from liability. The defense is governance built before the agent runs: hard spend caps, brand-safety constraints, audit logs, and a clear escalation path.
The third is the quiet erosion of judgment. If agents handle audience discovery, inventory sourcing, and optimization, the buyer's feel for the market, the instinct for which inventory is overpriced or which segment is decaying, can fade through disuse. That instinct is exactly what catches an agent when it is confidently wrong. A team that has outsourced the judgment may not have it when the agent fails.
For an enterprise weighing agentic media buying, the useful stance is neither early-adopter rush nor refusal. Treat an agent as a capable new operator who is fast and tireless and occasionally, expensively wrong. Hire it for the repetitive work where speed compounds and mistakes are cheap to reverse, not for the brief, the creative, the budget authority, or the final word. Demand that it log its decisions and expose its fees, set the spend caps and safety rules before it runs, and keep a person who still understands the market between the agent and the money. The shift from automated to autonomous is real and accelerating, and whether it serves a marketer or merely outpaces them depends on how carefully the guardrails are drawn before the agent is switched on. This is where an implementation partner like Perform Digital earns its place: not in picking the flashiest agent, but in drawing that control line correctly and building the governance that makes autonomy safe.
Council summary
This post argues that agentic media buying is a real break from Performance Max style automation: automation optimizes a campaign a human structured, while an agent does the structuring itself, moving a decision-maker rather than adding a tool. The review verified the load-bearing figures against primary sources, including PubMatic's AgenticOS launch, the Butler/Till and Geloso campaign results, the AdCP and AAMP standards timeline, and the Davis and Gilbert legal analysis. Two claims were corrected for precision: the 82 percent figure is a buy-side cost cut, not a generic supply-chain saving, and PubMatic's 80 percent growth applies to its emerging-revenue category that includes AI products, not an AI revenue line on its own. An unverifiable Duluth Trading quote was cut. The takeaway is concrete: hire an agent for repetitive, reversible work where speed compounds and mistakes are cheap, keep humans on the brief, creative, budget, and final approval, and draw the governance rules before the agent runs.
Comments