Agentic buying and the end of the ad manager
Agentic buying ends the ad manager as an interface, not as a job. The work moves to writing constraints and auditing decisions you no longer make by hand.
The ad manager, meaning the console where a human sets bids and budgets and chooses targeting, is on its way out as the primary way UA gets done. Moloco's help centre recorded at the end of April that app metadata editing in its Ads Manager is now restricted to internal staff. AppLovin opened Axon Ads Manager to self-serve e-commerce advertisers globally in late June, but the self-serve surface is thin by design because the model does the work. Unity's Vector takes an objective and a budget and returns results. On the buyer's side, AppsFlyer's State of Gaming for Marketers report in January already noted AI assistants in use for UA reporting, and most serious growth teams are now building the step after that: agents that act on the reporting.
The thesis is straightforward and, I expect, contested: agentic buying ends the ad manager as an interface but not the media buyer as a job. The job moves from operating the console to writing the constraints the agent operates within and auditing what it did. The teams that get this wrong won't be the ones that automated too much; they'll be the ones that automated the operating while keeping a headcount plan built for operators.
The mechanism: two models, one negotiation
Once your agent sets bids and the network's model accepts them, the negotiation is machine to machine. That sounds symmetric. It isn't, and the asymmetry is the whole point.
The network's model has seen every advertiser on its inventory, while yours has seen your account. It observes the auction directly; yours observes what the network reports. It is optimising its own yield, and yours is optimising your stated objective, which it will pursue literally, including into places you'd never have gone by hand. In a machine-to-machine negotiation the party with more information and a clearer objective wins. That party isn't the buyer.
This is why the console mattered more than it appeared to. A human in the console applied judgement the objective function did not capture: scepticism about a sudden CPI drop, a memory of what happened the last time a network offered cheap volume in a new geo, a sense that a creative was winning for the wrong reason. Automate the console and that judgement has to be written down or it disappears.
The second-order effect: the objective becomes the strategy
The effect most coverage misses is that an agent's objective function becomes the studio's de facto strategy whether anyone intended that or not.
Tell an agent to hit a day-seven ROAS target and it will find the cheapest path to that number. That may mean concentrating spend in a single network whose reporting is generous, pulling budget from the creative tests that feed next quarter's winners, or shifting into geos where it can hit the target on thin volume. Each of those is a strategic decision a growth lead would normally make explicitly; the agent makes them implicitly every hour and reports success.
There is a related trade-off in skills. A team that stops operating the console stops building the intuition that operating it produced. In three years the people writing constraints will be the people who used to set bids by hand. In eight years they will not be, and the constraints will be written by people who have never felt a campaign go wrong under their fingers. That is a real cost. The honest answer is that most studios will accept it for the efficiency, so the governance has to work without the intuition.
A framework: the constraint budget
Treat constraints as a budget you spend deliberately, not a list you grow reactively. An agent with too few constraints pursues the objective into places you would never have gone, and one with too many is a slow operator wearing a new badge.
Three layers have held up in practice.
Hard limits the agent cannot cross without a human: total daily spend, maximum share of spend on any one network, minimum share reserved for creative testing, excluded geos and placements.
Soft targets the agent optimises within those limits, meaning ROAS or cost per qualifying event at the horizon you choose, with the horizon set longer than any network's own optimisation window.
Tripwires that pause and escalate rather than adjust. CPI moving more than an illustrative thirty percent in a day is one; a new creative taking more than a set share of spend within forty-eight hours is another; a network's reported conversions diverging from your MMP or first-party count by more than a set margin is the one that catches the most.
The reserved creative testing share is the constraint most often left out and the one that matters most. Without it, the agent will starve exploration to feed exploitation, and the account will look excellent for two quarters before it runs out of things to scale.
A decision rule for autonomy
Not every decision should sit at the same autonomy level. A rule that has served well: grant autonomy in proportion to how reversible a decision is and in inverse proportion to how much money it moves.
Reversible and small decisions, such as bid adjustments within a network or pausing an underperforming variant, run fully autonomous and get reviewed weekly in aggregate. Reversible and large ones, such as shifting budget between networks, run autonomous within the hard limits with a same-day notification. Anything hard to reverse at any size (opening a new network, entering a new geo, changing the objective or the horizon) the agent proposes and a human approves.
Build the audit cadence around that third bucket and the tripwires. The first two will mostly be fine; reading them is where an operator's week disappears.
What the job looks like now
The media buyer of this decade writes specifications, checks the network's numbers against numbers the studio owns, and decides what the agent is not allowed to learn on its own. That is a smaller team than a console-operating one and a more senior one. The hiring mistake to avoid is keeping the junior operator roles and adding a specification role on top. The operator roles are what the agent replaced, and a studio that keeps them will find its people supervising a machine that no longer needs supervising while nobody writes down the constraints that do.
