Commissioning vs Steering

Wizard to patron

Wharton professor Ethan Mollick describes the shift in his own working vocabulary. With earlier models, working with AI felt like being a wizard: "you chant the spell and something happens," with the human casting and the model filling in detail under close direction. With a new class of model capable of sustained, multi-agent autonomous work, he is no longer sure he is the wizard: "I am closer to a patron. I describe what I want, I pay for it, and I judge the result." At the far end of the same progression, he says, "Fable is closer to a whole studio, where I am the client who signs off on the final work without ever setting foot on the floor."¹

The work shifts from process to outcome. Building an isochrone map and a nine-and-a-half-hour research tool he calls Concord, Mollick gave an ambitious instruction and a couple of minor pieces of feedback, while the model made "hundreds of little choices," in his words, without him understanding them or having a chance to weigh in along the way.¹ The model spins up its own research, coding, and verification subagents, and what comes back is finished rather than in progress.

The load-bearing claim: the black box may be permanent

The comforting assumption is that better interfaces will eventually restore visibility and mid-stream steering. Mollick resists that reading directly: "it is also possible that the opposite is true, that the more capable the model, the less there is for a human to meaningfully do, and the black box is the price of the power."¹ You can still steer, in the sense that instructions are followed remarkably well and the more ambitious the instruction the better the result tends to be, but supervision of process is replaced by judgment of outcome rather than restored by a friendlier interface.

The same shift, from the capability side

Scott Wu, co-founder and CEO of Cognition, arrives at the identical inflection point independently, naming the rungs by the unit of delegation rather than by the human's posture. Citing METR's research showing that unassisted AI task length has been doubling every couple of months, he treats the extrapolation as a first-principles question rather than a pattern match: seconds of delegated work is a command, hours is a task, and years is a mission, an agent studying a field and running its own experiments for months at a stretch.² Missions worth handing over, in his telling, include a societal problem to understand and coordinate on, a game design to iterate on end to end, or a novel line of materials-science inquiry the agent pursues on its own timeline.

The convergence between the two accounts is the real signal. Both agree that the scarce input flips from execution to specification once the time horizon crosses some threshold, and both place that crossing at multi-hour autonomy, one from the vantage of a user watching the process disappear, the other from the vantage of a builder watching the horizon extend.

The organizational echo

Jack Dorsey has described a company-scale version of the same shift, where humans increasingly sit at the edge of an organization's model, aligning it toward customer outcomes rather than directing each step of the work themselves. Commissioning, on that reading, is the individual-scale instance of the same edge-of-the-model alignment problem that a company built around AI has to solve organizationally.

The cost

The obvious tradeoff is that a subtle conceptual error becomes far harder to catch mid-stream if the process is never observed, so verification moves almost entirely to the finished output, raising the stakes of any failure that only shows up after a long autonomous run.