In 2024, AI agents came up in six of the stories we collected at Advertising Perspectives. In 2025, they came up in 127, and 2026 had passed 45 by spring.
We’ve written before about how every new technology in advertising follows the same curve, and agents are clearly on the steep part of it. But there’s a difference worth taking seriously. Most hype cycles ask brands to go somewhere new. Agents go where the work already is, into the buying, selling and planning that thousands of people do by hand every day. And unlike the metaverse, this one arrived with people arguing about standards almost immediately.
Here’s what the archive shows agents actually doing, and what we think is worth getting ready for.
Three kinds of agents
It helps to separate three things the trade press tends to lump together.
Agents that buy media. Software that takes a brief, plans a campaign, negotiates with sellers and places the buy, with a person approving the important steps. In January 2026, MediaPost reported that Yahoo’s DSP was advancing agentic buying models, and Beet TV’s headline was that AI agents were starting to buy TV ads for each other.
Agents that sell media. The same idea on the other side of the table. In October 2025, Digiday reported that The Sun, the U.K. tabloid, was building an AI agent for its programmatic business. This month, AdExchanger covered a company trying to build agentic infrastructure for CTV.
Agents that shop. This is the one with the biggest long-term consequences, and it isn’t about ad tech at all. In May 2025, MediaPost reported that Mastercard and Visa were building AI shopping agents. By October it was arguing that agents acting as brand influencers would collapse the consumer buying journey. If a consumer’s assistant does the research, compares the options and places the order, a lot of the advertising aimed at that consumer has to be aimed at the assistant instead.
A fourth thread runs alongside these: the conversations themselves becoming ad inventory. In January 2026, Digiday reported marketers trying to keep up as OpenAI geared up to launch ChatGPT ads, while another Digiday piece argued that the ads pivot may come too little, too late. By February, AdExchanger was writing about what happens when conversation becomes inventory and the arrival of the chatbot ad platform.
The fight underneath: standards
The detail that makes us take agents seriously isn’t any single product launch. It’s how quickly the industry started arguing about plumbing.
In October 2025, a group of supply-side companies including Yahoo and PubMatic formed a coalition to launch an agentic ad protocol, the Ad Context Protocol, which AdExchanger described as an attempt to make sense of agentic ad demand. Two weeks later, MediaPost reported that agentic ad creation would be tested through the Model Context Protocol, the general-purpose standard for connecting AI assistants to software, which AdExchanger later called the “universal adapter” for AI in advertising. In November, the IAB Tech Lab released its first framework for agentic ad buying standards, and in January it followed with a roadmap for scaling agentic AI.
We’ve seen this before. Programmatic only became a real market once there was a common way for buyers and sellers to talk to each other, and the fights over that standard shaped who held power for the next decade. The same thing is happening here, faster. Whoever defines how agents describe a campaign, an audience and a price will shape what agents can do, and for whom.
It’s also worth noting that the early warnings are already in. In November 2025, an AdExchanger contributor wrote why your first project built on the new protocol is probably going to fail. That’s not a reason to wait. It’s a reason to start small.
What agents need that most companies don’t have
Here’s the part that tends to get lost in the excitement. An agent is only as good as what it can see and what it’s allowed to do. That turns most of the agent conversation into a data conversation.
Clean, structured information about what you’re selling. A seller’s agent can’t negotiate on an audience it can’t describe. If your inventory, audience segments and pricing live in spreadsheets and people’s heads, an agent has nothing to work with.
Clear permissions. An agent acting for a buyer needs to know its budget, its brand rules and what it can commit to without asking. An agent working with customer data needs to know what each customer has consented to. These aren’t new requirements. Agents just remove the human who used to quietly fill in the gaps.
A customer relationship it can trust. WPP Media’s leaders put this well in August 2025: agentic commerce will redefine shopping, but trust and data come first. Epsilon made a related point this month, that loyalty has to leave the corner before agents rewrite commerce. If a consumer’s assistant is going to pick your brand, it needs a reason, and the strongest reason is a relationship the customer already values.
What we’d do now
Agents are early. Most of what’s announced today will change, and some of it will disappear. But the preparation that makes agents useful is the same preparation that makes everything else work better, so very little of it is wasted.
Get your own data in order first. Audiences that are defined, current and consented. Inventory and pricing that a system can read. This is the unglamorous work that every previous wave of ad tech also required, and it’s still where most companies are weakest.
Start with a narrow, internal use. Before letting an agent buy or sell anything, let it answer questions. Which audiences do we have? What did this campaign deliver? Which segments fit this advertiser? Internal questions are lower risk and teach you quickly where your data is thin.
Keep a person accountable for every decision that spends money. Agents should make people faster, not replace the judgment about what a brand is willing to do and pay. Every serious framework so far assumes a human approves the important steps, and that’s the right default.
Watch the standards, and participate if you can. The protocols being argued about now will decide which agents can talk to which systems. If your business depends on being bought or sold, you want a say in how that conversation is structured.
Think about the assistant as an audience. If consumers increasingly ask an AI to research and buy for them, your brand needs to be legible to that AI: accurate product information, clear terms, a reputation that shows up in the sources it reads. That’s closer to search optimization than to display advertising, and the brands that figure it out early will have an advantage.
The curve on agents will flatten, as every curve in our archive has. When it does, we expect the survivors to look less like a revolution and more like a large amount of ordinary work done faster, by companies that did their homework on data. That’s not a bad outcome. It’s how every lasting change in this industry has actually happened.
Every headline in this piece links to the original story. Browse the archive’s coverage of AI agents, agentic buying and ChatGPT ads, or read our look at a decade of ad tech hype cycles.
Kevin Reynolds is the founder of Advertising Perspectives, MadTech Consulting and data4orce.ai. He is a marketing technology architect and C-level advisor, with a focus on Salesforce and data-driven advertising.