The Macroeconomics of an Agent-Augmented Economy

Table of Contents

• Why Agents Are a Macro Question, Not Just an IT One
• The Production Function Gets a New Input
• Growth: Output Without Headcount
• Inflation and the Deflationary Pull of Cheap Cognition
• The Labor Share Problem
• Where the Productivity Gains Land
• The Measurement Trap
• What Central Banks and Treasuries Will Have to Relearn
• Insights Most People Overlook
• References

Why Agents Are a Macro Question, Not Just an IT One

Most coverage of agentic AI-as-a-service treats it as a procurement decision. You buy an agent the way you buy a SaaS seat, except now you pay per task completed or per outcome delivered rather than per login. That framing is fine for a CFO comparing vendors. It badly understates what happens when this becomes the dominant way cognitive work gets done across an entire economy.

The reason is simple. Labor is roughly 60 percent of value added in most developed economies. Anything that changes the cost, supply, and substitutability of labor on that scale is a macroeconomic event, not a line item. When the marginal cost of a competent knowledge worker drops toward the cost of compute and electricity, you are not optimizing a workflow. You are altering one of the two primary factors of production that every growth model since Solow has been built on.

That is the lens this article takes. Not "should my company use agents," but "what does an economy look like when a large and growing share of cognitive output comes from a near-marginal-cost workforce that nobody has to hire, train, or pay benefits to." The honest answer is that economists do not fully know yet, because we have no clean historical analog. But the structure of the question is clear enough to reason about carefully.

The Production Function Gets a New Input

Start with the classic setup. Output is a function of capital and labor, plus a productivity term that captures everything else, technology, organization, know-how. For a century, that "everything else" term, total factor productivity, was where technological progress hid. Computers, electricity, and the assembly line all showed up there.

Agents are different in a way that matters for the math. They are not just a multiplier on existing labor. They are closer to a new factor that is substitutable for labor across a widening set of tasks, and that is produced by capital (data centers, model training, inference infrastructure). In effect, you can convert capital directly into labor-equivalent units of cognitive output, on demand, at a price that keeps falling.

This is the structural feature that breaks intuition. In standard models capital and labor are complements over the relevant range, more machines make workers more productive but you still need the workers. An agentic workforce is a partial substitute that scales with spending rather than with population. There is no labor supply curve constraining it, no demographic ceiling, no commute, no sleep. The supply of agent-labor is bounded by chips and energy, both of which are expanding.

Andrew McAfee and Erik Brynjolfsson made the early version of this argument about digital technologies more than a decade ago, in their work on the "second machine age". Agents are the version of that thesis where the machines finally do the cognitive work, not just the arithmetic.

Growth: Output Without Headcount

The first visible macro effect is a loosening of the link between employment and output.

Historically, a growing economy meant a growing payroll. Okun's law, the rough relationship between GDP growth and unemployment, held because producing more stuff required more people. Recessions threw people out of work; recoveries hired them back. The two moved together tightly enough that policymakers treated employment as a real-time proxy for output.

An agent-augmented economy can grow output while employment stays flat or even falls, because additional cognitive capacity comes from spinning up more agent instances rather than from hiring. A 12-person company running fleets of agents can produce what used to take 200 people. We are already seeing early signals in software, customer support, and back-office finance, where revenue-per-employee figures at agent-leveraged firms look nothing like their predecessors'. The thesis of the small, agent-leveraged company, and its more aggressive cousin, the one-person company doing eight figures, is fundamentally a statement about this decoupling.

If that pattern generalizes, GDP growth can decouple from job growth in a way that has historically only happened briefly, and never structurally. That is genuinely new, and it scrambles a lot of policy machinery that assumes the two rise and fall together.

A caveat worth holding onto: this is the optimistic-for-output case. It assumes demand keeps up. An economy can produce vastly more, but if the productivity gains accrue to a narrow group and aggregate demand stalls because displaced workers have less to spend, you get more capacity chasing weaker consumption, a coordination problem, not a technology problem. That tension runs through everything below.

Inflation and the Deflationary Pull of Cheap Cognition

Here is the effect that is underappreciated relative to the job-loss headlines: agents are structurally deflationary for the price of services.

Think about what has driven the cost of living up over the past 40 years. It is not televisions or clothing, those got cheaper as manufacturing globalized. It is the labor-intensive services: healthcare, education, legal work, accounting, professional advice. Economists call the underlying mechanism Baumol's cost disease. Because a lawyer or a teacher could not be made dramatically more productive, their wages rose with the rest of the economy while their output per hour barely moved, so the price of their services climbed relentlessly.

Agents attack Baumol's cost disease directly. If a competent first draft of a contract, a tax return, a diagnostic summary, or a tutoring session can be produced at near-zero marginal cost, the productivity ceiling that drove those prices up is lifted. The likely result over a decade is real disinflation, possibly outright deflation, in exactly the service categories that have been the stubborn drivers of cost-of-living increases. The deflation of professional-services pricing is one of the most consequential second-order effects of the whole GaaS shift.

For central banks this is awkward. Their entire framework is calibrated to fight inflation that comes from a tight labor market and wage-price spirals. A persistent, technology-driven disinflation in services would push measured inflation below target for reasons that have nothing to do with weak demand, and responding to it with looser policy could inflate asset bubbles instead of helping anyone. This is the kind of supply-side disinflation that the 1990s tech boom hinted at, but potentially much larger and concentrated in services rather than goods.

The Labor Share Problem

The labor share, the fraction of national income that goes to workers rather than to owners of capital, has been drifting down in most rich economies since around 1980. There is a long debate about why: globalization, weaker unions, the rise of intangible-heavy superstar firms. The Brookings Institution and others have documented the trend extensively in research on the declining labor share.

Agents look likely to accelerate it, possibly sharply. If output increasingly comes from capital-produced agent-labor, then by construction more of the income generated by that output flows to whoever owns the capital, the model providers, the infrastructure owners, and the firms that deploy agents at scale. The "wage bill" for a unit of cognitive output shrinks toward an inference bill plus a margin.

This is the distributional core of the whole agent-economy debate, and it is where the macro story turns political. A more productive economy is unambiguously good in aggregate. But if the gains flow disproportionately to capital, you can have rising GDP, falling consumer prices for services, and a squeezed middle class with stagnant or falling labor income, all at once. Those three things are not contradictory. They are arguably the central tension of an agent-augmented economy, and the question of who captures the productivity gains does not have a market-determined answer, it is set by bargaining power, taxation, and policy.

There is a counter-mechanism worth naming. If agents make every individual worker dramatically more productive, augmenting rather than replacing, and if those productivity gains get bid into wages because skilled workers who can orchestrate agents become scarce and valuable, the labor share could hold up better than the pessimists fear. The historical record on this is mixed. Past general-purpose technologies eventually raised wages broadly, but with painful, decades-long lags during which specific cohorts of workers did badly.

Where the Productivity Gains Land

It helps to separate three different "pots" the productivity surplus can flow into, because confusing them produces most of the bad takes.

Lower Prices for Consumers

If agent-driven industries are competitive, much of the gain shows up as falling prices. Consumers capture it through cheaper legal help, cheaper software, cheaper customer service. This is the Baumol-reversal channel, and it is real, but it depends on competition. In concentrated markets, incumbents pocket the margin instead.

Higher Returns to Capital

If a handful of model providers and infrastructure owners sit at a chokepoint, they capture a large share as economic rent. The structure of the GaaS market, how many providers survive the consolidation endgame, directly determines how much of the surplus becomes durable profit versus competed-away savings. A market with five viable frontier providers distributes very differently from one with two.

Higher Wages for Complementary Workers

Some workers become far more valuable because they direct, audit, and combine agent output. The "agent boss" managing a fleet, the domain expert whose judgment the agent cannot replicate, the person who carries the human premium in services where clients pay specifically for a human. These people capture gains as higher wages.

Which pot fills up is not predetermined by the technology. It is determined by market structure, regulation, and tax policy. That is the single most important thing to understand about agent macroeconomics: the technology sets the size of the surplus, but institutions decide where it goes.

The Measurement Trap

There is a real risk that the official statistics badly mismeasure all of this, in both directions.

On one side, GDP may understate the gains. A lot of agent-driven value is consumer surplus, services that are now free or near-free that used to cost money. When something gets cheaper or becomes free, measured GDP can actually fall even as welfare rises, because GDP counts spending, not value received. The free-software and free-information economy already created this gap; agents widen it.

On the other side, there is the productivity paradox. The original version, Robert Solow's quip that "you can see the computer age everywhere but in the productivity statistics," took years to resolve as firms reorganized around the new tools. We may be in the same lag now with agents, visible everywhere, not yet in the aggregate productivity numbers, because reorganizing an entire firm around autonomous workflows takes time and the early deployments are clumsy. The MIT and Stanford economists who study this, including work referenced by the National Bureau of Economic Research on AI and productivity, generally expect a J-curve: things look disappointing, then inflect. Whether the inflection comes in three years or fifteen is the multi-trillion-dollar question, and it is exactly the productivity-paradox question for agents that deserves its own scrutiny.

What Central Banks and Treasuries Will Have to Relearn

Pull the threads together and a policy picture emerges that is genuinely uncomfortable for the institutions that run the macroeconomy.

Monetary policy loses two of its favorite signals. Employment stops being a clean proxy for output. And service-sector disinflation stops being a sign of weak demand. A central bank that keeps reading low inflation as "the economy needs stimulus" could fuel asset bubbles while the real economy is doing fine.

Fiscal policy faces an eroding base. Income tax and payroll tax systems are built on wages. If a growing share of value added flows to capital and to a small number of firms, the tax base narrows and shifts in ways the current system is poorly designed to capture. This is why serious proposals around taxing automation, capital, or compute keep resurfacing, not as ideology, but as plumbing. The design of those policy responses to agent-driven displacement will shape whether the productivity surplus gets recycled into broad prosperity or simply concentrates.

And the distributional politics get sharper. An economy that is richer in aggregate but more unequal in distribution is the classic recipe for backlash. The technology can deliver abundance and resentment simultaneously. Whether the agent-augmented economy is remembered as a golden age or a fracture depends far less on the models themselves than on the institutional choices made while they are being adopted.

None of this is a prediction of doom. The agent-augmented economy could be the thing that finally cracks the cost disease that has made healthcare and education unaffordable, that lets small teams build things only giants could before, that raises living standards broadly. But that outcome is a policy achievement, not a default. The macroeconomics says the surplus is real and large. It is conspicuously silent on who gets it.

Insights Most People Overlook

• The deflation is the bigger story than the job losses, and it cuts the other way politically. Everyone focuses on displacement, which is concentrated and visible. The diffuse benefit, collapsing prices in healthcare, law, education, the very services that have crushed middle-class budgets, is larger in aggregate and almost invisible in the discourse. A worker who loses a paralegal job but whose family's legal, medical, and tutoring costs fall by half is a genuinely ambiguous case, not an obvious loser. The net welfare math is harder than either side admits.

• "Marginal cost near zero" is a half-truth that will mislead a lot of business plans. The marginal cost of one more agent task is low, but the system cost, orchestration, error correction, liability, human oversight, the compute for reliability at scale, is not near zero. Economies built on the assumption of free cognition will discover, as the cloud era did, that the real bill is in the operational tail. The macro deflation is real but smaller than the naive model predicts.

• Aggregate demand, not production, is the binding constraint nobody is modeling. An economy can produce infinitely more cognitive output, but someone has to buy it. If the income to buy it is concentrating in capital, you can hit a wall where supply is unlimited and demand is the bottleneck, a glut of capability chasing shrinking purchasing power. This is the Henry Ford insight (pay workers enough to buy the cars) turned into a macro risk for the whole agent economy.

• The labor share could fall fastest in the richest economies and create a geographic plot twist. Because high-wage knowledge work is the most expensive to do with humans, it is the most profitable to automate first. That means the disruption may hit advanced economies' white-collar core before it touches lower-wage regions, partially inverting the offshoring pattern of the last 30 years, and possibly giving developing economies a window to leapfrog rather than be left behind.

• Mismeasurement could make the whole transition look like a failure right when it is succeeding. If GDP undercounts free and cheapened services while displaced wages show up clearly in the unemployment data, the statistics could show stagnation and pain at the exact moment real living standards are rising. Policymakers reacting to bad numbers could make genuinely bad decisions. The data lag is itself a macro risk.

References

• The Second Machine Age, Harvard Business Review
• The Decline of the U.S. Labor Share, Brookings Institution
• The Productivity J-Curve and AI, National Bureau of Economic Research