"Congress will not pass AI legislation this year." A senator said it last week without apology. The market did not move. Three days earlier I finished a re-audit of an AI-agent vault that had accumulated a nine-figure deposit base in under four months. Its documentation lists "regulatory clarity" as a key risk factor. There is no clarity. There is no statute. There is a vacuum, and the vacuum is the product.
I have spent ten years reading code that claims to be governed, and six months reading the filings of the agencies that claim to govern it. The two documents describe different universes. The first assumes rules will arrive. The second assumes they will not. Only one of them is right, and the market is pricing the wrong one.
The regulatory calendar is not mysterious. The White House issued an executive order on AI in October 2023. It carried reporting thresholds for large compute clusters, safety testing obligations for frontier models, and a mandate for agencies to publish guidance. It carried no legislative backing. An executive order is a variable, not a constant. The next administration can set it to zero.
The European Union took the other road. The AI Act classifies systems by risk tier. High-risk systems carry mandatory data governance, technical documentation, human oversight, logging, and conformity assessment. It applies to any provider serving the EU market, regardless of domicile. This is the same structure the bloc used for privacy under GDPR and for exchange operations under MiCA. Each time, the same mechanism repeated: market access conditional on regulatory conformity. Each time, non-EU firms paid the cost of adopting European rules rather than maintaining two systems.
That mechanism behaves like a tariff the exporter pays and the importer collects. The rent is collected in compliance, certification, and audit fees, all of which accrue to European assessors.
In the United States the alternative to federal statute is a patchwork. California writes one rule. New York writes another. The FTC asserts authority over unfair or deceptive practices. The SEC asserts authority over anything it can call a security. Courts fill whatever is left. Fifty jurisdictions, three agencies, and no single certification surface.
That is the landscape. Now the part the crypto market is not reading carefully.
The AI-agent sector sits precisely at the seam where two under-regulated systems meet. Agents that hold private keys, execute payments, and query price oracles fall under no clear regime in the United States. They are not securities in any obvious sense. They are not banks. They are not AI systems under a statute that does not exist. The absence is not neutral. It is a licensing regime with a zero fee and an undefined liability cap.
In 2025 I audited a protocol enabling autonomous wallets to act on oracle data. The feed validation logic looked like this:
function updatePrice(bytes calldata payload) external {
(uint256 price, ) = abi.decode(payload, (uint256, uint256));
oraclePrice = price; // no signer recovery, no nonce, no timestamp
}No signer recovery. No nonce. No freshness check. Any address with gas could write a price. The vulnerable surface was small, roughly forty lines, which is exactly why it survived three prior reviews. I ran close to 10,000 simulated manipulation paths over four days. Just under 400 returned a profitable path. The protocol paused its launch after the report went public.
The code spoke, but the logic was a lie. The documentation described a decentralized oracle network. The implementation described an open write function.
Here is the part that matters for policy. A regulated operator, whether exchange, custodian, or bank, has an examiner. The examiner asks about signature verification. The examiner asks about freshness checks. A statutory regime does not prevent the bug. It creates a party whose job is to look for it.
Nothing in the current US posture creates that party for agent protocols. The FTC acts after consumer harm. The SEC acts after a token is sold. Courts act after a plaintiff appears. All three are reactive. All three are slow. None of them audits a function before it is deployed.
Compliance is a fixed cost with a regressive structure. A conformity assessment under the European framework does not cost less because a firm is small. It costs the same. This produces a predictable distributional outcome: incumbents absorb it as a line item, startups die of it. Trust is a variable you cannot hardcode, and the cost of proving trust scales with the size of the entity, not the risk of the system. The result is not safety. It is consolidation dressed as safety.
There is one place where American policy strength has not decayed, and it is worth being precise about it. Compute. Chip export controls run through the Commerce Department. They are executive instruments. They do not require a statute. If the question is who can throttle the physical supply of training capacity, the answer remains Washington. The vacuum is not uniform. The United States has ceded the software and agent layer while retaining the hardware layer. That asymmetry is a policy choice, whether or not anyone intended it.
The AI Act regulates systems and providers. It regulates the model. It does not directly regulate an autonomous wallet that calls a model and signs a transaction. The EU crypto instrument, MiCA, regulates the wallet's service provider, if one exists. An agent with no legal person behind it falls into the gap between two regimes, in the same way the US agent falls into the gap between none.
So the Brussels effect will not arrive at the agent layer automatically. It will arrive through market access: the moment an agentic protocol needs a euro-denominated ramp, a licensed counterparty, or an exchange listing, a European condition attaches. That is the transmission channel. Not the statute. The on-ramp.
For valuations, the vacuum cuts both ways. Near-term cash flow improves, with fewer compliance hires, faster shipping, broader data use. The discount rate does not. Policy uncertainty is a pricing factor, and litigation is unbounded in a way that administrative rules are not. Copyright suits, negligence claims, and algorithmic discrimination cases will set the effective standard, one docket at a time. Data does not lie, but it does not care whether the rule governing a model arrives from a regulator or from a plaintiff.
The bulls are not wrong about everything, and it is worth conceding the strong version of their case. Legislative inertia is not identical to abdication. The United States still possesses the most aggressive extraterritorial enforcement apparatus in the world, spanning sanctions designation, secondary liability, and deferred prosecution agreements, and those tools move faster than any parliament. A rule written by enforcement action adapts in months. A rule written by legislature adapts in years, if at all.
There is a real argument that European speed is a liability. Drafted before the current generation of agentic systems existed, the AI Act's model-level obligations may be obsolete on arrival. Statutes are durable because they are slow. That durability is only valuable if the object of regulation stops moving. This one does not.
But the concession has a limit. Enforcement-driven rules are unpredictable, and unpredictability is itself a cost. A startup cannot price a subpoena. A fund cannot diligence a standard that will be announced retroactively. The European framework is expensive, but its expense is knowable in advance. The American framework is cheap until it is not, and the moment of expense is chosen by the prosecutor.
The interesting question is not whether Congress legislates. It is what happens to the first agent protocol that loses funds to an oracle write a signature check would have stopped. My working prediction: the first binding rule for on-chain agents arrives not from a statute but from a courtroom, or from a European exchange deciding that market access requires a certified signer. Watch the divergence in how protocols implement freshness and signer recovery over the next two quarters. That gap is the leading indicator.
They built a palace on a fault line. The fault is not in the code. It is in the assumption that someone else was checking it.
