Unpacking Agentic Advertising

On 15th Oct, 2025, a coalition including Scope3, Yahoo, PubMatic, and Magnite launched the Ad Context Protocol (AdCP). Within weeks of the announcement, the IAB Tech Lab released the Agentic RTB Framework (ARTF).

The core question: Should we build from scratch… or extend what works?

Follow along: We're going to trace a campaign from a brand manager's brief to a live media buy, watching both the buy side and the sell side at every step.

Every time a webpage loads, dozens of advertisers bid to show you an ad. The whole process is done in under 200ms. Blink and you miss it. This is Real-Time Bidding.

OpenRTB processes >600 billion bid requests per day. It was built to answer one question: what is this single impression worth, right now? It answers it brilliantly.

But every RTB auction is stateless and scope-limited, which means that it can't negotiate deals, adapt creative types, or shift budget when one channel outperforms.

RTB is optimized for speed and precision, not big picture strategy.

Before AI agents can transact, they need a way to interact with the outside world.

In Nov 2024, Anthropic released the Model Context Protocol (MCP). Often called the USB-C of AI, MCP gives AI systems to query legacy technologies by giving them a common tongue. Within months, OpenAI, Google, and Microsoft adopted it.

This brings new possibilities for advertising. If AI agents can freely interact to check inventory levels, negotiate price, or supply creative assets to a publisher, you have the foundation for something the industry had never built before.

Now, let’s say that Acme Outdoor's buyer agent sends out a campaign brief: "outdoor recreation, 25-45, CTV, Q2.” A publisher’s seller agent responds with: “2 million impressions, $28 CPM, pre-roll and mid-roll”.

The buyer agent makes a counter-offer. They go back and forth on the specifics of the deal. When both parties agree, the buyer agent executes the buy, pushes the creative, and starts pulling the campaign performance data.

Now, the important thing to note that is this is still a direct deal between the brand and the media entity, which means that it cuts out a lot of the intermediaries and the infamous “adtech tax” that they levy on every open exchange transaction.

A month after AdCP was announced, IAB Tech Lab fired back with ARTF.

Today, when an RTB auction runs, the bid request bounces between data centers. The SSP calls an identity vendor in Virginia. A fraud detection service in Dublin. A contextual engine in Oregon. Each hop adds latency. A round trip can take ~600-800 milliseconds.

ARTF eliminates the hops by containerizing those services inside the host's own data center. No network calls. The identity agent, the fraud agent, the contextual agent all operate on the same server, communicating via gRPC instead of HTTP.

Total processing time? ~100ms. That's an 80% reduction. The agents don't replace the auction. Each one returns a targeted edit instead of bouncing the bid around.

The industry now has (at least) two competing frameworks.

AdCP's philosophy: build a new language for AI agents from the ground up. Brian O'Kelley frames it as a shift from "expense-focused transactional thinking" to "investment-oriented strategic planning."

IAB Tech Lab's answer is AAMP, a framework that layers agent capabilities onto existing infrastructure: OpenRTB, ads.txt, sellers.json, etc. Anthony Katsur's argument: "When agents orchestrate complex workflows, ambiguity becomes catastrophic."

Then there's the walled gardens who are developing their own agentic protocols and are in no hurry to throw their weight behind any of the open standards.

Back to AdCP. For agents to trade, they need to read each other's identity cards.

On the buy side, brands publish brand.json on their domain. It declares the brand's name, logos, color guidelines, and a pointer to the brand's identity agent. When a creative agent needs to build an ad for Acme Outdoor, it pulls the specs from this file. No PDFs. No emails. No "can you resend the logo in SVG?"

On the sell side, publishers host adagents.json (which agents are authorized to sell my inventory) and sellers.json (which accounts can transact on my behalf). Together, they work like an ID badge system. A buyer agent checks both files before negotiating and the seller agent confirms authorization in the other direction.

Bidirectional verification prevents anyone from selling inventory they don't own.

Everything so far still ends with a human seeing a banner, a video, or hearing an audio spot. But there's one format that breaks the mold entirely.

AdCP includes a protocol called Sponsored Intelligence. When you ask your AI assistant about hiking gear, with your permission, a brand's agent can joins the conversation. Acme Outdoor's agent can then answer your questions about trail shoes, show comparison charts, and hand you off to a checkout flow. The conversation is the ad.

It's early (nobody is running this at scale yet). But it's the clearest signal of where agentic advertising diverges from everything that came before.

"Agentic advertising replaces human media buyers."

It doesn't. Agentic systems handle the mechanical parts of media buying (sending briefs, comparing proposals, pushing creative, pulling reports) but every protocol in this space includes governance layers requiring human approval. The goal is to collapse the 40-60% intermediary tax, not eliminate human judgment.

"AdCP and ARTF are competing to replace RTB."

They don’t. AdCP operates at the campaign level (strategy, negotiation, creative, measurement) while ARTF optimizes the auction itself (faster enrichment through co-located containers). RTB still handles the impression-level execution.

"Only big brands and publishers will use this."

The machine-readable identity files (brand.json, adagents.json, sellers.json) are deliberately designed to be as simple as ads.txt. A small publisher can host an adagents.json file just as easily as a major media company. The protocols are open-source and free to implement.