---
source_url: https://marquard.org/posts/agentic-buy-sell-loop/
title: "The agentic buy/sell loop just closed. What is the PM for now?"
date: 2026-06-15
summary: "Over the past month both ends of the ad transaction went agent-native — buying agents on Meta, Google, and TikTok’s MCP servers, selling agents on Scope3’s open AdCP. Once agents transact against agents on an open protocol, the defensible layer stops being the interface or the inventory and becomes the measured outcome the loop optimizes against — the one layer that doesn’t have an agent yet, and where the PM job relocates."
tags:
  - "ad tech"
word_count: 1207
reading_time_minutes: 6
---


Over the past month, both ends of the advertising transaction went agent-native — and neither did it inside a walled garden.

On the demand side, Meta, Google, and TikTok each [stood up MCP servers](https://digiday.com/marketing/tiktok-launches-mcp-server-to-let-ai-agents-run-campaigns/) so an assistant like Claude or ChatGPT can run a media buyer's campaigns across all three directly — set them up, adjust them, pull the numbers — with no human clicking through three separate ad managers. On the supply side, InMobi stood up what's been called one of the first [agentic sellers](https://www.emarketer.com/content/inmobi-sell-side-ai-agent-ad-tech-scope3) operating at real scale — wiring its exclusive CTV inventory into the Ad Context Protocol (AdCP), the open agentic-transaction standard Scope3 is building, so AI buying agents can transact against it with no salesperson in the seat.

Hold those two side by side. An agent on the buy side, an agent on the sell side, clearing against each other on a shared open standard. That is a closed loop. The thing the industry has called "agentic" for two years as a slide-deck adjective became a working transaction with software on both ends.

So the question for anyone running product in this market is no longer "are agents coming." It's the harder one: when both the buyer and the seller are agents, what is the human PM actually for?

## The loop closed from both ends

This wasn't one launch, and it wasn't one week. Demand moved first, supply followed weeks later, and only with that last piece in place does the whole loop come into view.

On the supply side, the pieces landed in early June. PubMatic shipped [Decision Fabric](https://pubmatic.com/blog/decision-fabric-how-we-put-partner-intelligence-inside-the-auction/), which puts partner AI decision models *inside* the auction instead of bolting them on after the bid — the SSP becomes a place where buying logic executes, not just a pipe that carries it. Then InMobi exposed that same premium, exclusive CTV inventory to autonomous buyers over AdCP — one of the first sell-side deployments to do it at scale. Two different moves toward the same thing: a sell side an agent can negotiate with.

The demand side had moved first, back in mid-May. The big-tech MCP servers hand the buying interface to whatever assistant the marketer already lives in. OpenAI's performance ads went [live inside ChatGPT](https://www.pymnts.com/artificial-intelligence-2/2026/openai-begins-offering-cost-per-click-ad-campaigns-on-chatgpt/) — the buy surface itself is now a consumer AI product. The media buyer's seat is dissolving into the assistant.

The connective tissue is the part that matters most, and it's the part most of the coverage skips: it's an *open* protocol. [AdCP](https://digiday.com/media-buying/wtf-ad-context-protocol/) comes out of Scope3, the agentic-ad company run by Brian O'Kelley — the AppNexus founder — with a founding coalition that includes Yahoo and PubMatic; he's deliberately pushed it toward independent governance, standing up a nonprofit so no single platform owns it. It's the same move MCP made on the model side: one open standard any agent can wire into, not another walled garden. The loop didn't close inside one company's product; it closed on rails the whole market can ride.

## The defensible layer isn't where the budget is looking

Most of the agentic-ads commentary has this backwards. Once both sides are agents transacting on an open protocol, you find what's defensible by elimination.

It isn't the buying interface. Claude or Codex or ChatGPT or whatever the marketer already has open will own that soon. Any ad platform betting its moat on a slick campaign UI is defending ground that just got conceded.

It isn't access to the inventory either. Aggregating access is what a DSP has always sold — but a DSP is a proprietary toll road: you integrate with it, you route your spend through it, and that position is the whole moat. An open protocol commoditizes the same access at the standard layer, where nobody owns it — a buying agent reaches the inventory directly instead of renting a lane through someone's stack. That's the difference AdCP makes: access stops being a product a middleman can charge a margin on and becomes something nobody owns. If that sounds like OpenRTB, it's a cousin, not a copy. OpenRTB standardized the sub-second impression auction — a rigid schema for bidding on inventory someone already packaged. [AdCP runs a layer up](https://ppc.land/ad-industry-veterans-explain-why-adcp-isnt-competing-with-real-time-bidding/): an agent discovers inventory against a brief, negotiates audiences, outcomes, and deal terms, and settles a whole campaign — the back-and-forth a fixed bid request was never built to carry. OpenRTB was a protocol for bidding; AdCP is a protocol for negotiating.

What's left, when the interface and the inventory both flatten, is the one thing the loop is wired *to*: the measured outcome it optimizes against. Whoever defines what a good outcome is — and can prove it actually happened — owns the layer the agents can't commoditize, because the agents are optimizing toward that definition. They don't write it.

## The outcome gap is the failure mode to design against

There's already a name for what goes wrong when nobody owns that layer. Call it the ["outcome gap"](https://aijourn.com/with-agentic-ai-brands-can-close-the-outcome-gap-or-make-it-worse/): the agentic stack turns every dashboard green — bids placed, budget spent, impressions delivered, the agent reporting success at every step — while the business result it was supposed to drive stays flat.

Automating the loop without owning the outcome definition doesn't close that gap. It widens it, because now the machine is confidently hitting a target nobody checked. Two agents can transact flawlessly all day, every metric green, and still optimize toward a proxy that has quietly come unbolted from revenue. Speed and automation make a misdefined outcome *more* dangerous, not less — the system pursues it faster and reports back that everything is fine.

## So what is the PM for?

The same thing the role was always for, with the parts that used to fill the day stripped away. When the agents run the campaign, the human's job collapses onto the work the agents can't do: write the spec, define the outcome, build the eval that proves the outcome is real, and set the guardrails for what the agents are allowed to do while no one is watching.

I [made this case from the hiring chair last week](/posts/ai-pm-job-description/) — that the trait worth screening for is whether someone defines "done" before they build anything. The agentic loop is why that's now the whole job and not a nice-to-have. Running the campaign — now the part with an agent — was ad ops' job, not the PM's, anyway. Defining the outcome and proving it happened is the part with no agent, and that has always been product's.

The loop closing isn't a threat to the PM job in ad tech. It's a spotlight on it. With execution automated and access commoditized, the one layer left worth owning is the definition of a good outcome and the proof it happened — and that's always been product's job. So when you look at a company shipping "agentic advertising" this year, the question worth asking is which layer it's building. Plenty are shipping the plumbing — the protocol connectors, the agent interfaces, the auction hooks. Far fewer are building the thing the plumbing is supposed to deliver. By the end of next year, that's the line that separates the companies that mattered in this shift from the ones that only ever sold the plumbing.