• by Maarten Albarda , Featured Contributor, 3 hours ago

The industry is charging headfirst into agentic advertising. I attended an Association of National Advertisers seminar this week, optimistically called “The Agent Wars.”

The  Interactive Advertising Bureau Tech Lab shared its evolving approach for Model Context Protocol (MCP) and Agent-to-Agent (A2A) workflows.

On paper, it looks brilliant. You give a business goal to a Campaign Strategy Agent. That agent collaborates with specialist sub-agents using the A2A protocol. Those sub-agents connect to ad servers, first-party data, and measurement platforms using MCP. They execute, optimize, and learn. Frictionless automation at scale and speed.

Humans? Optional, mostly, and positioned at the front end (“press start”) and end (“review result”).

Except there is a small problem. We already know large language models hallucinate. We know bad data creates bad outputs, and most marketers have terrible data taxonomy. Now we are chaining these models together, human centipede style, into autonomous loops.

What happens when my buying agent talks to a selling agent built by a developer with very different deliverables? How do I know that second agent isn't making up inventory numbers? What if the underlying model updated overnight, changing how my agent interprets risk?

This is the digital equivalent of the game of telephone.

My agent asks your agent for high-intent audiences. Your agent queries a publisher agent. That publisher agent checks with a measurement agent. If the third agent in the chain uses a rogue metric, my strategy agent reallocates real money based on fiction.

And because this happens in milliseconds inside an automated loop, nobody notices. That is, until the weekly budget report shows and the money has been spent.

I remember when automated bidding rules went rogue and blew through monthly budgets in 20 minutes. Or when the automated trading systems on Wall Street caused a real crash of P&G stock. Now imagine that, but with autonomous systems negotiating deal terms with each other on your budgets.

This isn’t an improvement. This is a chain of digital blind spots.

To make agentic advertising work, we cannot just trust the code. Governance is going to take time, money, and specialized skills. You need people who understand how these connected systems pass data, where the failure points live, and how to verify outputs.

If you cannot audit why Agent A made a decision after talking to Agent B, you are quickly flying blind. You've just outsourced your advertising to a sequence of black boxes you don't control.

Here’s what you probably should do before handing your media budget over to autonomous agents:

Require human-readable audit trails. Every A2A handoff needs a log. If an agent shifts budget or changes targeting based on a partner agent's signal, that decision logic must be stored, and you must be notified.

Do not let agents commit budgets autonomously without caps and rules. Any deal, bid increase, or strategy pivot over a specific dollar amount must require a human click to approve.

Verify the data sources and uses first. An agent is only as good as the database it hooks into via MCP. Clean up your internal taxonomy and first-party data pipelines, or your agent will just make bad decisions faster.

Agentic automation is here, but strategy still requires judgment. Don't trade accountability for speed.