One of these days, your media budget will be negotiated between two agents
Epicflare publishes with Open Garden a white paper on agentic AI and media buying, written for those who build agents. I read it from the advertiser's side: the negotiation of media buys is about to be automated end to end, and what will matter then is what your agent knows about you.
We are on the verge of fully automated negotiation of media buys. Your agent on one side, the media owner's agent on the other, and nobody left on the phone on a Friday evening to close a deal. That is what I take away from the white paper that Raphaël Ambit, founder of Epicflare, published on 14 October in Open Garden, Nicolas Jaimes's media. The guide is written for the agencies, adtech companies and sales houses that build agents; I am summarising it from the other side of the table, the advertiser's.
First the vocabulary, because everything gets mixed up. MCP is a connector: Ambit calls it "a sort of API 2.0, consumed by AI rather than by hand-written code". With it, an agent can read a platform or act in it. AdCP and AAMP are something else: open protocols, the first carried by a consortium around Scope3 with publishers and DSPs, the second by the IAB Tech Lab, with a shared vocabulary so that buyer agents and seller agents can talk to each other directly: find inventory, negotiate a deal ID, set up a campaign. MCP connects, the protocol organises the transaction. And as soon as an agent has to transact with the outside world, you go through the protocol. Ambit compares it to OpenRTB: nobody created their own version of the bidding protocol, everyone ended up speaking the same language. I think it will go faster this time.
Gartner has put a number on the first step: in August, the firm predicted that more than 70% of global ad spend would flow by 2028 through self-serve platforms where AI materially shapes buying, pricing and results. Nobody there is talking yet about agents negotiating with each other, but it is the same movement, one notch further.
The most useful point of the guide for an advertiser fits in one sentence: an agent placed in front of a protocol only brings something if it "injects expertise or data that do not exist in the protocol itself", business know-how, proprietary data and "decision rules specific to the agency or the advertiser". Otherwise, Ambit writes, it relays requests, adds latency and consumes tokens.
That is exactly mediagentic's bet. I have no platform and I will not invent a protocol. I write down what the buyer agent knows about my client before it negotiates anything: the plan, the caps, the media mix we want to hold, what we refuse, the conversion signals we send back to the platforms. Today my agents work through the official channels of Google, Meta, TikTok or Amazon; the day media owners open up their seller agents, the same rules will serve for the negotiation. Irreversible decisions, a budget we commit or a campaign we stop, remain validated by a human. And the price does not depend on the volume the agent negotiates, because I am paid neither by percentage nor by the hour.
One warning to finish. Since the start of the year, Meta Ads accounts have been suspended because unauthorised AI tools had been plugged into them, and most MCP connectors still only allow reading. If you connect an assistant to your campaigns, go through the platform's official access, and keep someone to validate what cannot be undone. The guide can be downloaded from the Open Garden article, it reads in an hour.