Define the term narrowly
An agent, usefully defined, is a system that takes a goal, decides on a sequence of actions, executes them against real systems, and observes the result. Not a chatbot, and not a rules engine with better copy.
By that definition, agent deployments in media operations are real but far narrower than the marketing suggests.
Running unattended
Three categories reliably run without a human watching. Monitoring and alerting agents that watch pacing, delivery, and cost per outcome and escalate anomalies with context. Reporting agents that assemble, normalize, and narrate performance data on a schedule. And QA agents that read a live campaign against its brief and flag mismatches.
All three share a property: the worst case of a mistake is a false alarm.
Human in the loop
Budget reallocation agents, which propose moves and execute inside guardrails a human set. Creative generation, where output is reviewed before it runs. Bid adjustment within bounded ranges. These work well and are widely deployed, but the guardrails are doing a lot of the work.
Remove the bounds and results degrade quickly — not because the model is stupid, but because the objective it optimizes is never quite the business objective.
Rolled back
Fully autonomous campaign launch, in most shops that tried it. Autonomous audience strategy. Client-facing agents that answer performance questions directly. The failure mode in each case was confident wrongness at a moment nobody was watching.
The rollbacks are underreported, which is why the deployment picture looks rosier than it is.
What this means practically
Agents have made the routine layer of campaign operations dramatically cheaper to run correctly. That is a large, real change, and it is the reason execution no longer needs to cost what an agency charges for it.
What still requires people: deciding what to sell, to whom, at what price, and knowing when a number is lying to you.