The Planner’s Temptation
The first version of the AI economy will look perfectly reasonable. It will arrive as an internal productivity initiative, not as an announcement that markets are ending.
A company will connect an agent system to documents, tickets, dashboards, customer records, code, policies, calendars, and workflows. The system will retrieve knowledge, draft decisions, assign work, and improve processes. Executives will value the control, security teams the containment, and finance the leverage. No one needs to identify as a central planner to build a highly centralized machine.
AI makes centralization look clean: one operating layer and one context bank, with every improvement feeding back into a corporate intelligence system. At first this is simply management with better tools. The twentieth-century managerial corporation already moved substantial coordination out of open markets and into administrative systems, a shift Alfred Chandler described as the rise of the “visible hand” (The Visible Hand). AI gives that hand more memory and speed.
The pattern can scale. First, companies centralize intelligence internally. Then platforms coordinate across companies. Then industries standardize around shared agentic infrastructure. Eventually, the state starts to look like the natural highest layer, the ultimate monopoly coordinator. The crude version of this fear is that AI makes techno-communism tempting. The more precise version is this: if productive intelligence is trapped inside institutions, the economy will drift toward larger and larger centers of agentic control.
Job displacement is only part of the risk. The individual may also cease to be the basic unit of economic learning. If people cannot own and carry a meaningful share of their AI-enabled capability, the value will accumulate wherever the agents live: inside companies, platforms, and governments.
What Markets Actually Move
Markets do not work merely because people sell labor hours. They work because people carry generalized capability from one specific situation to the next. A marketer learns how to diagnose a broken funnel in one company and brings that judgment to another. A software engineer learns architectural taste from one codebase and applies it to a different product. A sales leader learns how buyers stall, posture, panic, and commit, then carries that pattern recognition into a new market.
This exchange has always been messy. A worker leaves with knowledge the company helped create, while the company keeps systems the worker helped build. Some of it is documented or protected by contract, and much of it lives in a brain that cannot be audited in any practical sense.
Economists have long had names for pieces of this. Human capital is one of them (Human Capital). Friedrich Hayek’s classic argument about dispersed knowledge is another: the knowledge required to coordinate an economy is not given to any one mind or institution, but scattered across people and local circumstances (The Use of Knowledge in Society). Ronald Coase explained that firms exist partly because markets have coordination costs, which means the boundary between firm and market is always a practical bargain, not a sacred line (The Nature of the Firm).
AI does not repeal these truths. It forces the labor market to say out loud what it has always depended on: people do not arrive at a company empty, and they do not leave empty. The productive economy is built on the movement of generalized learning through specific institutional contexts.
AI makes the boundary harder to ignore because some of that learning can now live outside the person.
The Human+Agent Pair
In a future labor market, a worker may bring more than a résumé and portfolio. She may also bring a portable context layer containing workflows, patterns, prompts, evaluation criteria, domain heuristics, operating principles, and agent skills accumulated over years.
“Skill” here should be understood broadly. It might resemble the way current agent platforms talk about tools, instructions, and reusable capabilities, whether in agent frameworks, Claude-style skills, or coding assistants such as Codex (OpenAI Agents SDK, Claude Skills, OpenAI Codex). But the deeper idea is not tied to any product. A skill is useful context that helps an agent act well. It is the difference between a generic assistant and one that knows how a particular professional thinks, checks, prioritizes, and ships.
Imagine a programmatic advertising specialist. In today’s market, she brings experience. In an agent-mediated market, she might also bring a portable system that knows how she evaluates campaign performance, structures creative tests, diagnoses conversion leaks, allocates budgets under uncertainty, and decides when an apparent winner is only statistical fog wearing a little crown. This system should not contain a previous employer’s customer lists, internal dashboards, unreleased strategy, or proprietary data. But it may contain generalized judgment about how to do the work.
The productive unit is the human+agent pair, not the agent alone. The person still decides what matters, interprets ambiguity, senses politics, weighs tradeoffs, and recognizes when a technically correct answer is professionally foolish. The agent carries part of the accumulated method. Together they can be more capable than a worker with a laptop without becoming a company-owned artificial employee.
This extends the argument that professionals should own portable AI augmentation rather than surrender every improvement to whichever institution currently employs them (SWE: Own Your Own AI). If intelligence keeps improving quickly, the ownership layer around human+AI capability becomes one of the central institutional questions of the next decade (2028 Intelligence Explosion).
The Boundary Problem
The difficult question is what belongs to the worker and what belongs to the company.
Companies own specialized context: internal data, customer records, proprietary strategy, codebases, pricing models, private workflows, compliance rules, product plans, system access, and competitive information. Individuals should not be able to scoop this material into a private agent, resign on Friday, and use it on Monday to help a competitor eat the old employer’s lunch.
Individuals, however, own generalized context: transferable methods, domain judgment, reusable patterns, personal workflows, learned heuristics, professional taste, and ways of thinking abstracted from specific employers. A growth marketer should not leave with a company’s proprietary acquisition model. But she should be able to leave with improved judgment about why paid campaigns fail, how funnel metrics lie, and how to design tests that do not flatter the person who designed them.
The line is not clean today and will not become clean with AI. Labor markets already rely on imperfect distinctions between employee learning and trade secrets, professional growth and confidential information, tacit knowledge and theft. AI makes those distinctions explicit, machine-readable, and politically unavoidable.
Companies will fear this, and they should. A portable context layer can store, retrieve, generalize, and apply knowledge at scale. If badly governed, it could operationalize leakage rather than merely remember it. The old world had leakage through brains, lunch conversations, slide decks, and suspiciously familiar strategies. The new world could have leakage through executable know-how.
There is one potential advantage: unlike a brain, a context layer can in principle be inspected and constrained.
Context Escrow
A workable answer is more likely to involve negotiated portability with enforceable boundaries than total openness.
One possible model is a third-party context host, or context escrow. The individual keeps generalized context there while the company keeps specialized context inside its own systems. During hiring, the parties define what can flow between them. The company’s agents access the worker’s context through a control layer. Both sides can improve through the work without the company automatically absorbing the worker’s entire portable system.
This is not a complete solution. It is a direction of institutional design. The agreement might define what is portable, restricted, locked, time-limited, or competitor-sensitive. Certain company-specific learnings might remain locked for a period after departure. Certain generalized skills might be usable at non-competing firms but restricted for named competitors. Some abstractions might be allowed, while raw data, customer-specific patterns, and proprietary workflows remain prohibited.
The idea sounds unfamiliar because the current system hides the same problem inside people. Contracts, confidentiality obligations, invention assignments, trade-secret law, norms, and litigation already police the boundary between employee learning and company ownership. The result is imperfect, expensive, and often unfair.
The agentic version needs stronger law, better defaults, and sharper limits. It must not become DRM for workers, where companies lock up a person’s future earning power by claiming ownership over every improved habit. It must not create context landlords who become the new monopolists of professional identity. It must not turn hiring into an invasive scan of a person’s intellectual life. The existence of bad versions is an argument for designing the category before the worst version arrives.
Why Companies Should Want It
For companies, the main attraction is faster onboarding and execution.
A firm that hires only people has to wait for them to translate experience into local action. A firm that hires only internal agents risks training its future on its past. But a firm that can safely hire human+agent pairs imports generalized capability in a form that can begin working almost immediately. It is not merely acquiring labor. It is acquiring a portable, tested, adaptable slice of professional method.
AI systems grounded only in internal context can become intellectually inbred. They optimize current practice and inherit the company’s blind spots. That is useful until current practice is the problem; then the agent repeats yesterday’s assumptions with greater speed and confidence.
Portable worker context creates controlled outside air. A company hiring a programmatic advertising specialist is hiring a human+agent pair that has seen other funnels, constraints, failure modes, and testing cultures. The specialist brings generalized skills C and D. The company has specialized systems Y and Z. In the work, C and D improve. Y and Z improve. The worker may leave with a new generalized skill E. The company may keep a new specialized capability V. Both sides compound.
At its best, this is exchange rather than extraction: a repeatable way to turn specific work into generalized learning and apply generalized learning to new specific work. It will not produce a perfect division of value, because none exists now.
Context as Credential
Hiring relies on noisy signals. Résumés compress a career into a favorable story, interviews reward fluency in an artificial setting, and references vary in candor. Work samples help, but only to a point.
Portable context could become a better signal. Not because companies should get to inspect a candidate’s private system in full. That would be grotesque. The better model is mediated evaluation. A third party could compare the company’s needs against the candidate’s context without revealing the raw contents of either side. It could run simulations, generate match profiles, estimate onboarding speed, and identify likely contribution areas.
A company might learn that a candidate has strong generalized capability in bid optimization, conversion-funnel diagnosis, creative testing loops, or budget allocation under uncertainty. The candidate would not have to expose every workflow. The company would not have to reveal its internal strategy. The signal would be compatibility, not naked access.
Done well, this could make labor markets more efficient and less theatrical. Done badly, it could become another surveillance funnel. The difference will lie in ownership, standards, and bargaining power. A context credential should increase a worker’s leverage, not strip-mine it.
The Market Needs Portable Intelligence
AI will make some firms much more productive. The open question is whether that productivity remains locked inside institutional walls or circulates through the economy.
Centralized systems optimize what they can see. They are good at coherence, consistency, enforcement, and scale. They are bad at surprise. They reduce variation by design. They prefer the measurable to the weird, the sanctioned to the emergent, the known workflow to the awkward new trick that looks wrong until it changes the game.
Markets need the opposite qualities too. They need variation, selection, failure, recombination, mobility, competition, and the occasional unreasonable person who refuses to optimize the local optimum. Portable individual context gives productive intelligence a way to move. It lets firms learn from the outside without simply stealing from the outside. It lets workers learn from companies without simply leaking the companies. It turns the economy into a larger search process.
The alternative is not neutral. If AI productivity accrues only to centralized institutions, the social logic bends toward dependency: wages from AI-rich firms for the few, platform access for the managed many, state redistribution for the displaced. Universal basic income or universal basic services may become necessary in some scenarios, but redistribution after value creation is not the same as participation in value creation.
A healthy AI economy needs more than a way to pay people after machines produce the surplus. It needs a way for people to remain economically generative. It needs an atomic mechanism of value creation and reward. That mechanism used to be the worker carrying skill in a brain. The next version may be the worker carrying skill across a human+agent system.
The Brain Was Always Portable
The centralized future is administratively attractive. It promises safety and efficiency while reducing the mess of markets, the friction of hiring, the ambiguity of work, and the leakage of learning.
But the mess is not a bug in the market economy. It is part of the engine. People learn in one place and apply the lesson somewhere else. Companies benefit from what workers bring, then lose some of what workers become. Workers benefit from what companies teach, then carry generalized ability into the wider world. The boundary has always been contested. It has also been indispensable.
AI does not abolish that bargain. It requires us to rebuild it with machinery attached.
The future should not be one giant agentic bureaucracy, nor a lawless bazaar where every company’s secrets bleed into every worker’s private model. The better path is harder and more interesting: individual-owned generalized context, company-owned specialized context, negotiated access, enforceable boundaries, third-party custody where useful, legal limits against abuse, and real mobility.
The economy already depends on workers carrying what they learn. The task now is to build institutions that let them continue doing so safely when part of that learning is stored in an agent system.
