everyone in adtech spent years going back and forth about cookies, clean rooms, retail media, CTV, and supply path optimization. Now it feels like the next big fight is shaping up to be AI agents.

not "AI helps write ad copy" or "AI summarizes reports." i mean actual agentic advertising workflows: buyer agents, seller agents, automated campaign setup, budget pacing, bid logic, creative testing, inventory discovery, measurement checks, and maybe eventually agent-to-agent negotiation between platforms.

On paper, this sounds like the natural next step for programmatic. Programmatic was supposed to automate media buying, but anyone who has worked in the ecosystem knows how much manual work still exists. Teams still spend hours moving numbers between platforms, checking discrepancies, rebuilding reports, troubleshooting tags, refreshing audiences, chasing screenshots, and asking why a campaign that looked fine yesterday suddenly went sideways.

So the pitch is obvious: let agents handle the repetitive execution layer while humans focus on strategy.

But the more interesting question is not whether AI agents can optimize campaigns. They probably can. The real question is: who gets to control the agent?

if a DSP builds the buyer agent, will it optimize for the advertiser's true outcome or for spend retained inside that platform? If a SSP builds a seller agent, will it optimize publisher yield transparently or just route demand toward preferred pipes? If a walled garden builds an agent, will it ever recommend moving budget outside its own walls?

That is where this gets messy.

Adtech already has a trust problem. Buyers complain about opaque fees, unclear supply paths, inflated reach, inconsistent attribution, made-for-advertising inventory, and platform-owned measurement. Publishers complain about take rates, auction dynamics, identity loss, and declining control. Now imagine putting an AI layer on top of that without clear auditability.

An agent that says "I optimized your campaign" is not enough. Optimized against what? Incremental conversions? Last-click ROAS? Viewability? Attention? Margin? Platform revenue? lowest CPM? highest win rate? Highest probability of getting the dashboard to look good?

If agentic advertising becomes real, I think the industry will need a few non-negotiables:

Clear agent identity: who built it, who controls it, and whose interests it represents.
Transparent objectives: what the agent is actually optimizing for.
Audit trails: what decisions it made, when, and why.
Permissioning: what data, inventory, and budget the agent is allowed to access.
Measurement independence: proof that performance is not only being graded by the same platform that executed the buy.
Human override: the ability to stop or adjust automation before it burns through budget.

There is a version of this future that is genuinely useful. A media buyer could say, "Find me high-quality open-web video supply against this audience, avoid MFA, cap household frequency, prioritize incremental reach, and explain every major budget shift." A publisher could say, "Package inventory in a way that protects yield, avoids channel conflict, and rejects demand that fails brand or privacy rules."

That would be powerful.

But there is also a darker version where every platform launches an "AI agent" that is basically a black-box sales rep with API access. It recommends more spend, hides complexity behind friendly summaries, and makes it even harder to understand where money went.

My guess: agentic advertising will not replace ad ops, traders, or yield teams overnight. But it may change what those jobs look like. The valuable people will be the ones who know how to set constraints, question outputs, debug bad recommendations, and understand incentives across the stack.

The worst mistake would be treating AI agents as neutral. They are not neutral. They will reflect the incentives, data access, and business model of whoever built them.

So maybe the real question for adtech is not "Can agents buy media?"

It is: "Can we build agents that are transparent enough for buyers and publishers to trust?"

Curious how people here are thinking about this. Are agentic workflows actually showing up in your stack yet, or is this still mostly conference-deck material?