The Tokenization of AI Agents: Notes From the Border of Two Fields

Writing the Future of Jobs Reports for the past four years has put me on edge. The labor economics side of my work keeps pointing toward a structural shift that existing financial instruments weren’t designed to handle. Can tokenization be a possible mechanism? I spent time taking it seriously as someone looking for frameworks that labor economics may not yet have.

What follows is an attempt to map where the two fields connect, where crypto’s tooling addresses real problems, and where I remain genuinely uncertain.

A note on what I’m analyzing: the unit of analysis throughout is the agent as a productive asset, a domain-specific AI system with persistent identity, attributable output, and a deployment feedback loop. The firm is the holder of that asset. The token is a claim on it. These are distinct layers, and the argument operates at all three, so it’s worth keeping them separate from the start.

Something Structurally New Has Happened to Labor

Software, patents, and recorded media already separated ownership from the worker. What’s new is different and more precise: a compressed approximation of expert cognitive procedure can now be replicated without the worker present, at near-zero marginal cost, across a bounded but commercially significant range of tasks.

Just to clarify: what is being replicated is not the full depth of expert judgment, including situated reasoning in genuinely novel cases, relational work, decisions that require reading context that isn’t in the data. What is replicated is a functional approximation, effective within the domains and task types represented in the training distribution. Outside those bounds, it degrades. The approximation is incomplete.

But incomplete replication is enough to break the negotiation floor, and this is where the labor economics needs to be made explicit. Sufficient replication reduces the scarcity of the cognitive output. Reduced scarcity weakens the worker’s outside options, the credible alternatives they can invoke when negotiating. Weaker outside options lower the reservation wage, the minimum the worker will accept. You do not need to fully replicate an expert to structurally alter their bargaining position. You need to replicate them well enough, often enough, for the tasks that constitute the core of the role.

39% of workers’ core skills are expected to change by 2030 across surveyed firms in the Future of Jobs Report 2025. What that number doesn’t capture is the ownership question underneath: when the procedure is approximated at scale, who holds the resulting asset? There is, unfortunately, no clean answer in existing labor economics, because the field was built on the assumption that cognitive procedure stays with the person.

The economic consequence is straightforward: firms that previously competed for expertise now face a portfolio question, own the capability, license it, or cede that market to whoever does. That is a fundamentally different decision than hiring, with different instruments, different time horizons, and different risk profiles.

Where Value Lands

The infrastructure layer of AI (i.e., compute and frontier models), is controlled by a small number of vertically integrated firms. If they absorb the agent layer too, there is no external market to structure and the rest of this argument is moot. This fork needs to be stated plainly because everything downstream depends on which branch holds.

Branch one: vertical integration dominates. OpenAI, Anthropic and Google deploy specialized agents directly. Enterprise customers buy access through existing cloud relationships. No independent builder sustains a performance advantage. No external agent market forms. Tokenization becomes a niche instrument for crypto-native use cases. As Vitalik Buterin has observed, the risk is that just as privacy-preserving technology was gaining mainstream adoption, a handful of cloud providers come to mediate all intelligent activity, and the financial architecture question and the sovereignty question collapsing into the same problem at different layers.

Branch two: domain specialization persists as a defensible advantage. Proprietary training data creates performance gaps that scale alone cannot close, and companies who design and deploy domain-specific models are able to move faster and answer these demands, as outlined in my podcast with Wendy this week (in Mandarin though). A legal research agent trained on ten years of firm-specific litigation outcomes knows something a general-purpose frontier model does not, and that gap proves durable at the task level.

The evidence for Branch two to become possible is specific and growing. Research published in Nature Communications across twelve biomedical NLP tasks found that fine-tuned domain-specific models outperformed frontier models in the majority of structured information extraction tasks. Benchmarking published by Together AI found a fine-tuned 27-billion parameter model outperforming a leading frontier model by 60% on a healthcare domain task at one-tenth to one-hundredth of the compute cost. Medical, legal, and scientific domain fine-tuning consistently produces 40-100% performance improvements on structured tasks.

That task-level performance gap is what makes the market structure claim possible. A durable performance advantage creates a defensible asset. A defensible asset supports independent value capture. Independent value capture is what makes an external token market meaningful. Without the performance gap, you have no asset worth tokenizing. With it, the question becomes which instrument best captures its value, and that is the question the rest of this piece addresses.

Everything that follows is conditional on branch two holding. If frontier scale eventually closes the domain gap, the argument collapses at this step.

Why Existing Instruments Fall Short, and What Tokenization Natively Provides

Assume branch two holds. Domain-specialized agents exist as a durable category. What is the right financial instrument for that market?

Before reaching for tokenization, it is worth being precise about why existing instruments fall short. Short answer: they fall short not because they are unsophisticated, but because each fails in a specific and structural way for this asset class.

Equity fails most visibly on governance, and the specificity objection is partly answerable. A critic will note that firms already use SPVs, tracking stocks, and spin-outs to isolate specific assets. Yes, equity can, with sufficient legal engineering, approximate the specificity that agent tokens provide natively. That’s true. But the governance problem remains even after isolation: an equity holder in a spin-out has no mechanism to govern what the underlying agent does, how it is retrained, or what use cases it is directed toward. You own the vehicle, not the wheel.

Royalties fail because they are static claims on an evolving asset. A royalty gives you a defined share of revenue from an agent as it exists today. But agents are continuously retrained. A royalty holder has no mechanism to govern retraining decisions: whether the domain shifts, whether a new version degrades performance in the original use case, whether the asset is effectively replaced while the revenue stream continues under the same contract. The static claim breaks precisely when the underlying asset changes, which in this asset class is the norm.

Employment contracts fail because they were designed for recurring labor, not permanent cognitive transfer. The contract expires when the worker leaves. The encoded approximation doesn’t, as outlined in my previous article.

What domain-specialized agents require is an instrument that 1) isolates a specific productive asset, 2) governs its evolution over time, 3) allows fractional ownership, and 4) enables a secondary market, without requiring bespoke legal structuring to approximate each feature separately. The pattern across the existing instruments, equity, royalties, and employment contracts, is the same: existing instruments address one or two of the requirements this asset class demands, but not all four simultaneously.

Tokenization is not the only conceivable instrument. But it is the one that natively combines all four features in a single structure. Equity and royalties can approximate them, but at significant coordination and legal cost, and without the governance right that distinguishes an actively managed productive asset from a passive revenue claim.

A token in this framework is a programmable, transferable claim on a defined share of an agent’s deployment revenue, with governance rights over retraining parameters. The revenue claim is the financial instrument. The governance right is what distinguishes it from a royalty and what resolves the hardest conceptual problem: if the agent is continuously retrained, what does the token represent? It represents whatever token holders decide it continues to be. A philosophical problem becomes a mechanism design problem with a designated decision-maker.

The Distributional Question Is the Most Important One

This is where I return to the general economics discussion, and where I think the crypto community has not fully grappled with what it has built.

As discussed in my article “Who Paid the Humans Who Fed the Machine?”, The workers who generated the training signal were compensated for a world where their procedure stayed with them. TIME’s investigation into OpenAI’s annotation workforce documented wages of $1.32 to $2 per hour for workers providing feedback that made frontier models commercially deployable. OpenAI paid their outsourcing partner $12.50 per hour for the same work. The gap between those two numbers is not a labor market inefficiency. It is a measurement failure: the contracts were priced for a world where cognitive procedure stayed with the person, not for a world where a compressed approximation of it would be permanently encoded and deployed at scale indefinitely.

Token allocation in AI projects went to founders and venture capital. That reflects who bore financial risk. The structural observation is different: the asset that was built partly from labor never priced for permanent transfer is now held entirely by people who were in the room when the tokens were designed. The people who weren’t in the room have no claim on what their contribution became.

Tokenization at least makes this a solvable financial problem rather than an invisible one. Contributors to supervised fine-tuning and RLHF have logged, timestamped, attributable sessions. Their contribution can be measured against training performance on specific tasks. They can receive vesting token tranches tied to deployment success, contributing now, vesting as the agent generates revenue, with acceleration if their contribution domain drives measurable performance gains. This mechanism is technically tractable at the fine-tuning and RLHF layer. It breaks down at pre-training, where the contribution surface is the entire internet and individual attribution is impossible.

The scope of the claim is therefore precise: the window to compensate the human training pipeline is the fine-tuning layer, before those roles are themselves automated. The Future of Jobs Report 2025 projects that AI will create 11 million new roles by 2030 while displacing 9 million and the roles being displaced include the very data-entry clerk type of positions that currently feed the fine-tuning pipeline. After that window closes, the distributional question doesn’t get resolved. It becomes structurally invisible, because the mechanism that would make compensation possible disappears along with the roles themselves.

That window is open now. Not for long.

What I’ve Concluded, Tentatively

Coming from labor economics, the instinct is to ask whether the instruments in use match the underlying economic reality. What I find in AI agent markets is a genuine mismatch: a new class of productive asset with compounding yield, governance complexity, and a distributional problem baked into its creation, being handled with instruments designed for a world where cognitive procedure walked out with the worker.

One more honest caveat belongs here, and it connects back to the fork stated at the outset. If vertical integration wins, i.e., if the agent layer is absorbed by infrastructure players before an independent market has time to form, then the tokenization argument becomes a description of what could have been rather than what will be. The argument is not deterministic. It is conditional on branch two, and branch two is an empirical bet that domain specialization outpaces scale-driven homogenization. That bet may not pay off.

Tokenization is not a perfect fit for the problem even if it does. The markets will be thinner than advocates claim. The governance mechanisms solve some alignment problems and introduce new ones. The distributional fix is technically feasible but requires deliberate design choices that nobody is currently making.

But the conclusion is not “this might not work, but it’s interesting.” It is more precise: the alternatives fail structurally for reasons already named. Equity can isolate assets with legal engineering, but cannot govern them. Royalties generate revenue claims, but cannot govern evolving assets. Employment contracts were not designed for permanent cognitive transfer. Given those failure modes, an instrument that natively combines specificity, governance, composability, and secondary markets, even imperfectly, even in thin markets, is preferable to bespoke approximations of each feature assembled at high coordination cost.

Tokenization is not the answer. But, tokenization can be the instrument whose failure modes are worth engineering around, rather than ones that make the problem structurally unsolvable. That distinction matters, because the field is young enough that getting the framing right now determines what gets built and what gets built determines who, eventually, holds the tokens.