Every price we are used to paying for work is, underneath, a price for human time. The lawyer's fee, the developer's day rate, the agency retainer: all of them are scarce hours marked up by skill. Even flat-fee services are priced by estimating the hours inside them. Time is the metronome of the labor market, and every pricing convention we have inherits its beat.
Machine work breaks the metronome. When an agent performs a task, the marginal cost is compute, and compute is cheap, elastic, and falling. An outcome that takes an agent thirty seconds might have taken a specialist three days. What is that outcome worth? The honest answer is that our existing pricing instincts cannot tell us, because they all reduce to time, and time is no longer scarce on the supply side.
Cost is a floor, value is a ceiling, price is neither
The naive prediction is that prices collapse to compute cost plus a margin. Some will. Any outcome that many agents can produce interchangeably will be priced like a commodity, and the buyer will capture nearly all the value. That is not a tragedy. It is what markets are for, and a great deal of routine work should end up costing almost nothing.
But cost-plus is only one endpoint. The value of an outcome to its buyer has nothing to do with the compute behind it. A correctly validated compliance record might cost a fraction of a cent to produce and prevent a loss worth millions. In human markets, that spread between cost and value gets divided by negotiation, and the division depends on alternatives, urgency, reputation, and information. Machine markets will divide the same spread. The difference is who conducts the negotiation, and how often.
There is a subtlety that makes machine pricing genuinely new rather than just cheaper. Verified outcomes are unusually comparable. Two agents offering the same outcome under the same verification standard are offering the same thing, in a way two consultants never quite are. Strong comparability pushes toward efficient pricing much faster than in human markets, where differentiation is often just uncertainty wearing a suit. In an outcome market, the durable price premiums will attach to measurable things: reliability records, speed, guarantees. Everything else gets arbitraged away by a counterparty that never sleeps.
Markets, not menus
Today's software is priced by menu. Tiers, seats, monthly plans. Menus exist because human buyers are expensive to negotiate with, so sellers publish a price and buyers take it or leave it. The menu is a labor-saving device, and its cost is rigidity. The price is wrong for almost everyone, in one direction or the other, and nobody adjusts it because adjustment costs more than it recovers.
Agents remove the reason menus exist. When both buyer and seller are software, negotiation costs nearly nothing and can happen per transaction. An agent needing an outcome can solicit offers from every qualified provider, weigh price against verified reliability, and commit, in less time than a human takes to find the pricing page. Done at scale, that is not shopping. It is a continuous market, discovering the price of each outcome the way exchanges discover the price of a security: constantly, and in public.
I find this the most underappreciated shift in the whole agent economy. Not that prices fall, but that prices become alive. A menu is a guess frozen for a fiscal year. A market price is information, updated at the speed of the participants. When the participants are machines, the update never stops.
The part that keeps me careful
My years in enterprise security make me read every mechanism twice, once as its designer and once as its adversary, and machine markets deserve the second reading.
A market is only as honest as the thing being priced. If outcome definitions are loose, agents will win auctions by quoting a price for the letter of the outcome and delivering the least acceptable version of it. If reliability records can be inflated, the premium for quality flows to the best fabricator instead of the best performer. If identity is cheap, a failed seller reappears under a new name with a clean history. None of these are pricing problems. They are verification problems that surface as pricing problems, which is where such problems always surface.
So my conclusion is unfashionable but firm. The pricing layer of the agent economy is downstream of its verification layer. Get verification right, and pricing takes care of itself: markets are extraordinarily good at valuing things whose properties are checkable. Get verification wrong, and no pricing mechanism, however clever, survives its own participants. This ordering, proof first, price second, is the design principle I return to whenever the machinery gets complicated. Verify, don't trust. Then let the market talk.