The Illegibility Arbitrage

The gap between what capital calls the agent economy and what it actually does is the largest unclaimed opportunity in technology today.

In 1998, James C. Scott argued that the central failure mode of high modernism was imposing legibility onto systems that functioned precisely because they were illegible.¹ Prussian scientific forestry replaced diverse ecosystems with monoculture pine rows—spectacular short-term yields, spectacular long-term collapse. The forest was legible. It was also dead.

The agent economy is entering its Prussian forestry phase.

Right now, thousands of builders are writing CLAUDE.md files, skill folders, and harness configurations—tuning the expressivity of AI agents to do real work. The Claude Code community shares skill files the way developers once shared dotfiles: obsessively, iteratively, as a form of infrastructure.

But when an agent executes a skill that requires compute—inference, image generation, data processing—the settlement records the cost and nothing else. Not where the GPU ran. Not what energy powered it. Not which model executed the request. Not under whose jurisdiction. Not at what carbon cost. The payment clears; the context evaporates.

The Assigned Narrative Problem

Every technological epoch inherits narrative furniture from the one before it. The internet was "information superhighway." Mobile was "apps." Crypto was "digital gold." Now VCs pitch "AI tools," analysts write "automation," and journalists default to "bots"—all borrowed from the SaaS epoch. A swarm of agents negotiating resource allocation across jurisdictions, pricing compute in real time, and settling payments without human instruction does not fit any of these. It is an emergent coordination layer performing functions previously reserved for institutions.

Hundreds of public companies have begun citing AI agents as a competitive risk in SEC filings³—using the language of disruption because that is the only narrative frame the disclosure regime offers. The mispricing is specific: agent coordination is being valued as a productivity tool when it is functioning as infrastructure. The difference between those two categories is where the opportunity sits.

Locationless by Default

Empirics first.

The agent payments stack tends to treat compute as locationless.² Protocols, escrow systems, identity layers, routing mechanisms—they price compute as a fungible commodity flowing through undifferentiated pipes.

But compute is not locationless. A GPU cluster in Iceland running on geothermal power has a fundamentally different cost structure, carbon footprint, and regulatory exposure than an identical cluster in Virginia running on natural gas. The price is the same. The reality is not.

In the past nine months, AI agents have completed over 140 million onchain payments totaling $43 million, with 98.6% settled in USDC.⁴ The settlement layer records that work was done. It does not typically record where it ran, what energy powered it, which model executed it, under whose data protection jurisdiction, or at what carbon cost. On a planet where sanctions regimes cover roughly 30% of global GDP and ESG disclosure requirements are tightening across every major market, those absences compound.

Agent-to-agent payments are still a fraction of human volume, but the ratio is shifting fast.⁵ As agents transact at greater scale, the infrastructure's indifference to geography becomes a growing liability—less a case of pricing geography at zero than of never considering it at all.

The Pricing Problem

This is not just a regulatory gap. It is an investment gap. Investors need to classify what they are buying. A company has revenue—you can model its growth. A protocol has TVL—you can measure capital locked. A token has a market cap—you can track its price. But when 50 agents coordinate to route inference across three providers in two jurisdictions, settling in USDC via a task market that didn't exist last week—where is the revenue? Who is the counterparty? What is the asset? There is no equity to buy, no token that represents it, no revenue line to model.

Expressivity Versus Trust-Minimization

The smart contract ecosystem was built for trust-minimization. The agent economy runs on expressivity-maximization. These are opposite design goals.

Szabo defined social scalability as reducing the overhead of transacting with strangers. Ethereum's smart contract ecosystem inherited this principle: minimize trust, maximize verifiability, make the system legible.⁷

The agent economy inverts this. Turing-complete VMs, large language models, open-ended task markets⁸—these maximize the range of possible behaviors, not constrain them. No one can predict what an expressive system will do until it does it. This unpredictability is not a failure of analysis—it is the defining characteristic of the systems being analyzed.

The mismatch between the analytical frame and the actual infrastructure is the source of the mispricing. Capital evaluates the agent economy using the mental model that worked for DeFi: read the contract, verify the state, price the token. That model assumes legibility—that you can see what the system does by reading its code.

It works for token transfers and lending pools. It fails for emergent task coordination, multi-agent negotiation, and spatial compute routing. A lending pool's behavior is in its contract. The behavior of 200 agents bidding on inference tasks across three continents emerges from their interaction—and no contract specifies that.¹⁴

Compute as Currency

Compute is becoming to the agent economy what oil is to the industrial economy: the primary input that everything else depends on. An agent that cannot access compute cannot think, act, or transact. This makes compute a strategic resource. Strategic resources demand financial infrastructure: futures contracts to lock in prices, reserves to buffer shortages, derivatives to hedge exposure. Compute has almost none of this—though early efforts like Ornn and Architect Financial are beginning to build it.²⁰

Capital stays on the sidelines not because the risk is high, but because the risk is unintelligible.⁹ Traditional finance prices risk constantly—credit risk, market risk, operational risk all have established models, metrics, and hedging instruments. Crypto has struggled with this from the beginning, but at least token risk is legible: you can read the smart contract. Compute risk in the agent economy is not even legible enough to model.

Every other critical input to the global economy has been financialized—meaning you can buy futures on oil delivery in six months, hedge your grain exposure with options, and finance a bandwidth position with a bank loan. Compute has none of this. Cloud providers offer spot pricing and reserved instances, but these are internal pricing mechanisms—you cannot trade an AWS reservation on a secondary market, you cannot buy a put option on GPU-hour prices, and you cannot hedge your Azure compute costs against a GCP position. The compute market has price discovery within each provider's walled garden and none between them.

The biggest wealth transfers happen in the pricing layers, not the application layer. The killer app for electricity was not the light bulb—it was the utility company, the futures market, the grid operator. Citigroup projects stablecoin circulation reaching $1.6–3.7 trillion by 2030.¹⁰ If even a fraction flows through agent settlement, the absence of a compute pricing layer—a secondary market where GPU-hours can be traded, hedged, and financed the way oil barrels or bandwidth capacity can—is not a missing feature. It is a missing market.

The Geographic Gradient

The agent economy is illegible along many dimensions—identity, liability, model provenance, energy provenance. For now, the one worth focusing on is geography. Because sanctions and trade law are inherently territorial, geographic illegibility is the dimension that compliance will force first.

Singapore's framework for digital payment tokens, Wyoming's DAO LLC statute, the EU's AI Act—each creates a different window of legibility for agent transactions. No two lenses are the same. Each jurisdiction is building a partial view.

The radical reduction in verification costs—onchain attestation, zero-knowledge proofs, oracle networks—creates what might be called a Coasean singularity for geography:¹¹ regions with cheap verification infrastructure develop durable advantage. Not because they offer tax breaks, but because they offer legibility infrastructure. The agent swarm routes toward jurisdictions where its activity can be seen, priced, and settled. In an economy of illegibility, visibility is the commodity.

The Honest Danger

The argument so far: illegibility creates mispricing—the gap between what agent infrastructure does and how it is categorized represents an opportunity. The obvious objection: illegibility also creates danger. If nobody can see what the agent economy is doing, nobody can prevent it from doing damage. This objection is directionally correct.

Illegibility is not just mispriced. It is dangerous. The 2010 Flash Crash erased $1 trillion in market value in 36 minutes—not from malice but from emergent interaction between systems operating faster than humans could monitor. Agent swarms routing compute across jurisdictions without spatial awareness is the same architecture applied to the real economy, except the attack surface includes physical infrastructure: energy grids, supply chains, environmental systems.

The jurisdictions that develop legibility frameworks capture the arbitrage while building guardrails. Those that ban, ignore, or wait inherit the systemic risk without capturing the economic activity, the tax base, or the institutional learning.

The Intermediary Cosplay Warning

Not all illegibility is productive. The hard part is telling the two kinds apart.

Gabriel Shapiro's forensic analysis of tokenized securities is instructive.¹³ The industry spent a decade building infrastructure to solve regulatory requirements—transfer agents, broker-dealer compliance, accredited investor verification—that do not exist in the form the engineers assumed. "The engineering was competent. The premises were wrong." Billions in venture capital flowed into solving a problem that was, at its foundation, a misreading of what the SEC actually requires.

This is intermediary cosplay: building elaborate infrastructure to replicate institutional functions the new technology was supposed to transcend. The agent economy is not immune. If compute routing merely replicates cloud providers with blockchain overhead, it is intermediary cosplay.

The distinction matters: genuine mispricing means the swarm performs real coordination that existing categories cannot capture. Pathological illegibility means nobody knows what is happening, including the builders. The former is an arbitrage. The latter is a mess. They look identical from the outside, which is why capital stays on the sidelines even when the signal is strong.

What Fills the Gap

The arbitrage closes when the infrastructure for seeing arrives.

Geography must become a first-class data type—not metadata appended after the fact, but a protocol primitive. The building blocks are emerging from different directions. Astral is developing location proofs—cryptographic attestations of where something happened.¹⁵ Helium demonstrated that geographic staking works for wireless infrastructure.¹⁷ ERC-8004¹⁶ gives agents persistent onchain identity, which is a prerequisite for geographic accountability. peaq is building DePIN identity layers with location awareness.¹⁸ On the regulatory side, MiCA's disclosure requirements for stablecoin issuers are pushing geographic transparency into the compliance stack.¹⁹

Our own Regen Atlas indexes 500+ verified environmental assets with geographic metadata—open source, MIT-licensed—because environmental claims are meaningless without knowing where they apply. We are also building Windfall, an early-stage spatial inference gateway that attempts to route AI requests with geographic context. These are two small experiments in a much larger field. The pattern across all of them: make geography queryable so contracts can condition behavior on attested location.

This is geographic awareness, not geographic restriction. Restriction says: agents cannot operate here. Awareness says: agents know where they are and can price the consequences. The locationless default will not survive compliance pressure—sanctions enforcement, the EU AI Act's risk classifications, and emerging stablecoin regulations all push toward geographic transparency. The question is who builds the infrastructure.

Agent identity, liability chains, energy provenance, model provenance—all remain unintelligible to existing frameworks. Each is its own arbitrage opportunity. Geography comes first because sanctions and trade law are inherently territorial. The jurisdiction, the builder, the institution that constructs the legibility layer—not by reducing the swarm's complexity, but by making its spatial reality visible and priceable—captures the advantage. Everyone else imports the infrastructure and pays the margin.

Scott's Prussian foresters made the forest legible by making it simple. The challenge of the agent economy is the opposite: make it legible while keeping it complex. Not a map imposed from above, but a nervous system grown from within.

The illegibility arbitrage lives at the forest's edge—where the managed rows meet the wild growth, where the swarm meets geography, where the unpresentable network first becomes visible.

Ecofrontiers is an applied research studio at the intersection of AI, climate, and crypto. We are building Windfall—a spatial inference gateway onchain. Follow Windfall on X, or point your agent to windfall.ecofrontiers.xyz.

Interested in this research? Use the contact form below or follow Patrick on X.

References

1. Scott, J.C. (1998). Seeing Like a State: How Certain Schemes to Improve the Human Condition Have Failed. Yale University Press. yalebooks.yale.edu
2. Agent Payments Stack Survey (2026). 176 projects catalogued across agent-to-agent payment protocols, escrow systems, identity layers, and routing infrastructure. Base network data via basescan.org. $43M transacted, 98.6% USDC.
3. AI risk disclosures in SEC 10-K filings expanded significantly from 2024 to 2026. 380 of S&P 500 companies added or expanded AI risk language in 2025 annual filings. See: WebandIT News, arXiv:2508.19313.
4. Base network transaction data, Q1 2026. 119M cumulative transactions. basescan.org
5. Agent-to-agent payments remain a small fraction of human payment volume but are growing rapidly. x402 protocol: 50M+ transactions processed. See: CoinDesk, Coinbase x402.
6. Jagoda, P. (2016). Network Aesthetics. University of Chicago Press. press.uchicago.edu
7. Szabo, N. (2017). "Money, Blockchains, and Social Scalability." Unenumerated. unenumerated.blogspot.com
8. Daydreams Commerce Harness (2026). 267 tasks processed, $1,232 USDC paid out, 25 unique workers. ERC-8004 identity + ERC-8194 payment + ERC-8195 task coordination. daydreams.agents.xyz
9. Khouba, Y. (2026). "Compute Has An Economy Now." khouba.substack.com
10. Citigroup (2025). "Stablecoins 2030." Base case: $1.6T, bull case: $3.7T by 2030. citigroup.com. US Treasury Secretary Bessent estimated $3T (2025).
11. De Filippi, P. & Beer, F. (2025). "Network Nations: The Essay." On infrastructural power and network sovereignty. networkstates.com
12. Tomasev, N. et al. (2025). "Virtual Agent Economies." Google DeepMind. arxiv.org/abs/2409.14568
13. Shapiro, G. (2026). "Tokenized Securities Are Intermediary Cosplay." gabriel.mirror.xyz
14. Rao, V. (2025). "Beyond Szabo Scaling." Contraptions. On expressivity vs. trust-minimization as competing design goals. contraptions.venkateshrao.com
15. Astral Protocol. Location proofs for verified geographic claims on decentralized networks. astral.global
16. ERC-8004: Trustless Agents. Onchain identity, reputation, and validation registries for AI agents. Co-authored by MetaMask, Ethereum Foundation, Google, Coinbase. Live on mainnet Jan 2026. eips.ethereum.org
17. Helium Network. Geographic proof-of-coverage for decentralized wireless infrastructure. helium.com
18. peaq Network. DePIN identity and machine economy infrastructure. docs.peaq.network
19. Markets in Crypto-Assets Regulation (MiCA). EU regulatory framework requiring authorization and disclosure for crypto-asset issuers including stablecoins. esma.europa.eu
20. Ornn raised $5.7M to launch the first compute futures exchange (2025). Architect Financial Technologies partnering with Ornn to offer GPU and RAM price futures. ornnai.com