Agentic Advertising Is Coming: When the Buyer and Seller Are Both Agents

Software is going to buy and sell media. Not “assist a trader,” not “optimize a campaign overnight” — actually plan, negotiate, transact, and report, with no human in the loop for the routine cases. The buyer will be an agent. The seller will be an agent. And the thing they trade — the ad — will increasingly land inside a conversation rather than on a page.

If you have been in ad tech for a while, that sentence should feel familiar in its bones. We have watched a version of this movie before. The interesting question is which parts are a rerun and which parts are genuinely new.

We have seen automation eat the insertion order before

A decade and change ago, buying media meant a salesperson, a phone, a PDF insertion order, and a fax-machine cadence that took days to clear a deal. Then programmatic happened. The insertion order got abstracted into a bid request, the negotiation got abstracted into an auction, and the whole thing started clearing in roughly a tenth of a second per impression. If you want the mechanics, we wrote a plain-English primer on real-time bidding — but the short version is that machines took over the transaction.

What programmatic did not change is who was on each end of the deal. There was still a human-defined campaign on the buy side and a human-defined inventory package on the sell side. The auction was automated; the intent on either side was authored by people and frozen into rules ahead of time. The machine executed a plan. It did not make one.

That is the line the current wave is crossing.

What is actually new: both ends can reason

The phrase “agentic advertising” went mainstream fast in 2025. Scope3 put it on the marquee with an agentic advertising platform in March. By Cannes Lions in June it was the dominant hallway conversation, and — tellingly — the question had already shifted from “what is this” to “okay, how do we actually do it.” When the industry stops debating the definition and starts arguing about implementation, something real is underneath.

Here is what is real underneath. Two standards quietly became load-bearing:

• MCP (the Model Context Protocol). Anthropic open-sourced it in late 2024, and through 2025 it became the de facto way to connect an agent to tools and data. The part that matters for advertising is that MCP is self-describing. An agent can ask a server what it can do and get a structured answer back. A buying agent does not need a hand-written integration for every supply source; it can discover capabilities at runtime.
• A2A (Agent2Agent). Google introduced it in April 2025 as a way for agents to talk directly to other agents rather than only to tools. That is the missing primitive for a negotiation. A buyer agent and a seller agent need a channel to haggle over, not just an API to call.

Put those together and you get the thing programmatic never had: a buy side that can form intent on the fly and a sell side that can respond with judgment, negotiating in something closer to natural language than a fixed bid schema. The plan and the execution collapse into the same step.

The transaction is also catching up to the rhetoric. In September 2025 OpenAI and Stripe shipped an agentic-commerce capability that lets an agent actually complete a purchase, and Google announced an agent-payments effort with a long list of partners. I will not overstate these — they are early. But “an agent that can transact” stops being a thought experiment the moment an agent can move money. Buying media is just another transaction.

The part nobody has solved: where does the ad render?

Here is the operator’s warning, and it is the same warning programmatic should have taught the industry the first time.

Standardizing the transaction is necessary and not remotely sufficient.

RTB solved how to clear a price for an impression. It did almost nothing about whether the resulting banner was any good, whether it belonged on the page, or whether anyone wanted it there. We spent the following decade paying for that gap in viewability fraud, banner blindness, and ad blockers. The auction was efficient. The experience was an afterthought, and the afterthought is what eventually capped the whole model’s value.

The agentic wave is sprinting toward the same cliff. Most of the energy right now is on the buying side — the protocols, the negotiation, the agent that plans a campaign and goes shopping. That side will get standardized; the incentives are too strong and the pieces (MCP, A2A, agent payments) are already on the table. I would bet that an agent-to-agent ad-buying standard is close to inevitable, even though none has launched publicly yet.

But none of that answers the question that actually decides whether agentic advertising is worth anything: when a buying agent wins, what does the user see, and where?

If the surface is an AI chat, the answer cannot be a banner. There is no page. There is a thread, a turn, a moment in a conversation. An ad there has to read like a message, not an interruption — which is a much higher bar than “the creative loaded.” It has to:

• Earn its place in the turn. The ad competes with the model’s own answer for the user’s attention and trust. If it does not match the live intent of the conversation, it is noise, and noise in a thread is far more expensive than noise on a page.
• Render natively, server-side, model-agnostic. Whatever model the user is talking to, the ad has to arrive as a first-class part of the response, brand-safety-filtered, in the format the surface actually uses.

This is the native-versus-interruptive distinction, except the stakes are higher because the medium is more intimate. We have written before about what native advertising really means and about how that plays out specifically inside answers when Perplexity started putting ads in AI answers. The throughline: in conversation, the placement layer is the product. Get it wrong and no amount of buy-side automation saves you.

Where Elo fits, honestly

Elo is not trying to be the agent-to-agent buying standard. That layer is going to get built, probably by several parties, and it should be open.

What we are building is the unglamorous, unsolved piece on the other end — the serving and placement layer for the conversational surface. Elo reads the live context of a chat, matches it to an advertiser offer, runs brand-safety filtering, and delivers the ad server-side as a native message in the thread, across web, iOS, and Android, regardless of the model underneath. When the buying agents arrive — and they are arriving — they still need a surface where the impression renders as something a person actually wants in that moment.

That is the half of agentic advertising that does not get solved by a protocol.

The auctions will go agentic. The transaction will standardize. None of it matters until the ad earns its turn.