AI Agents Are Now Buying Media. Is Programmatic Entering Its Next Era?

PubMatic built a product for AI agents, not human traders. A CTV campaign already came in with CPMs 18% below target.

A media trader's job has always been three verbs: plan, execute, optimise. PubMatic just built a product designed to let an AI agent do all three without waiting for a human to approve each step.

What AgenticOS does

PubMatic's AgenticOS is built for AI agents rather than human traders, letting an agent plan a campaign, execute the buy, and continuously optimise it against a target, all inside the same workflow a trading desk used to run manually across multiple tools. The pitch is speed: an agent can react to a shifting auction in the time it takes a human to open a dashboard.

The clearest evidence so far comes from a Connected TV campaign run for Telefónica by agency Havas, where PubMatic reported CPMs landing 18% below target. That is not a marginal efficiency gain. It is the kind of number that gets a CFO's attention faster than any explanation of how the technology works.

What this actually changes on a trading desk

For years, "programmatic optimisation" meant a human trader adjusting bids based on a dashboard that updated every few hours. An agentic system collapses that loop. It watches the auction, adjusts the bid, and reports the outcome, continuously, without a person in between deciding whether today's adjustment made sense.

That raises the obvious question every trading floor is now quietly asking: what is the human trader for, once the agent can plan, buy and optimise faster than a person can review the plan? The likely answer, at least for now, is judgment on the input side, choosing the target, setting the guardrails, and checking the agent's results against reality before green-lighting more budget.

The Ad Tribe angle

Programmatic advertising has always been sold as an efficiency story. AgenticOS is the version of that story where the person negotiating the buy is no longer a person. Whether that is progress or a new kind of risk depends entirely on whether the data feeding the agent is as good as the CPM number it produces, and that is a question the industry has not fully answered yet.