In 1976, Jensen and Meckling formalised what every manager already felt: delegation is expensive because the person you delegate to has their own interests. They called the gap "agency costs" — the monitoring, the incentive design, the residual loss when interests diverge anyway.
A century of management practice is machinery built against that problem. Bonuses. Equity. Performance reviews. Career ladders. Culture decks. All of it exists to make someone else's hands serve your intent.
Half the problem just vanished
An AI agent has no self-interest. No careerism, no shirking, no politics, no quiet quitting. The incentive problem — the half of delegation economics that consumed the most managerial energy for a hundred years — simply does not exist.
This is why delegating to agents feels so strange to experienced managers. The instincts they spent decades building — read the room, align the incentives, protect the relationship — find nothing to grip.
The other half doubled
What remains is information asymmetry, and it arrives inverted. With human delegation, the worker knows more about the work than the manager — that's the classic asymmetry. With agents, the output lands on your desk fluent, confident, and formatted beautifully whether it is right or wrong. The surface carries no signal.
Every agency cost that used to be spread across monitoring, incentives, and alignment now collapses into a single point: verification.
Incentive design was the skill of managing people. Verification design is the skill of managing agents.
What verification design looks like
Not reading everything — that just makes you the bottleneck again. Verification design means: a definition of done the output can be checked against. Spot-checks aimed where the work is most likely to fail. Approval gates sized to the stakes. And when output is wrong, diagnosing which part of the brief failed — context, scope, or definition of done — rather than concluding the technology isn't ready.
This is why "skim it, ship it" is the most expensive habit in the agent workforce, and why full autopilot deliberately scores below gated autonomy on the DQ assessment. The Review dimension isn't bureaucracy. It's where all the cost went.