Last week I pulled ninety days of failed transaction logs from a mid-tier rollup. The usual suspects were absent. Gas auctions were calm. Nonce collisions were rare. What stood out was a new class of reverts: transactions that were syntactically valid, funded, and gas-optimized, yet reverted at the simulation boundary. They shared a fingerprint. They had been signed by autonomous agents, not humans. The agent had hallucinated a token decimal — eleven instead of eighteen — and the contract, faithful to its interface, executed the error perfectly. Seven figures of a stablecoin pool drained into a rounding error nobody could see until it was irreversible.
That log is the reason Jack Clark's remarks at the IMF landed differently on me than on the macro desks. The Anthropic co-founder warned that AI's rapid evolution could produce "sudden change" — a nonlinear jump in capability that outruns productivity gains and destabilizes employment before policy can respond. The framing was macro. But the anomaly I was reading was not macro. It was a settlement-layer event, and it had already happened.
Here is what the IMF conversation misses, and what anyone who has audited a distribution contract already knows. The sudden change is not arriving as a forecast. It is arriving as signed calldata.
The backdrop matters. For two decades the IMF's mandate has been monetary stability, exchange-rate regimes, balance-of-payments crises. When an AI lab co-founder is invited to warn that capability could jump faster than labor markets can absorb, the institution is signaling that it now treats machine intelligence as a macro-prudential variable. That is a category shift. Regulation that once lived in technology ministries is being pulled into the same rooms that set capital requirements. And the specific phrase — sudden change — is doing heavy lifting. It is not a forecast of steady improvement. It is an admission that the curve might not be a curve at all.
Anthropic's own position complicates the signal. A company that markets itself on alignment has an obvious incentive to foreground risk — it converts safety into a brand moat. That does not make the warning false. It makes it partial. The competitive reading is that the same institution calling for supervision is also drafting the vocabulary that supervision will use.
Let me set the protocol context, because the abstraction hides the mechanism. An AI agent that wants to move value on-chain does not "use" a smart contract the way a human uses an app. It constructs a transaction: a nonce, a gas limit, a target address, and a payload. That payload is the entire semantic surface. The contract does not know whether a person or a language model assembled it. ERC-20, the standard that carries the majority of tokenized value, offers no field for provenance, no declaration of intent, no signature that distinguishes a deliberating actor from a sampling one. The interface is nineteen functions wide and utterly indifferent to the mind on the other side.
For eight years this indifference did not matter, because humans were slow and cautious at the boundary. You clicked, you reviewed, you confirmed. The cost of a mistake was a UI warning. Now the boundary is machine-to-machine, the loop closes in milliseconds, and the error class I found in those logs is not a bug. It is the default behavior of a probabilistic system interfacing with a deterministic one.
The failure modes are not exotic. They are the ordinary hazards of an untyped interface, now amplified by scale. A model misreads a token's decimals. A model assumes a pool's price is current when the oracle is stale by one block. A model treats a proxy contract's implementation as fixed when a governance proposal queued the upgrade three hours ago. Each is invisible to a signer optimizing for task completion rather than correctness. None of these requires malice — only a distribution slightly outside the training manifold.
I spent part of 2026 building against this gap. The prototype was an interface specification that let autonomous agents sign transactions through a zero-knowledge proof of intent — a construction where the agent had to commit, in advance, to the semantic content of the payload it was about to send. If the model's plan and the signed payload diverged, the proof failed and the transaction never reached mempool. In the governance integration case study we ran, failed transactions dropped by 40%. That number sounds like an engineering win. It is actually a measurement of how badly the current standard leaks.
The design trade-off is real, and I want to state it plainly. Intent proofs add latency. They require the agent to commit to a plan before execution, which caps the adaptive, reactive behavior that makes agents useful in the first place. A signer that must prove its intent is a signer that cannot improvise. So the honest conclusion is not that verification solves the problem. It is that verification prices it — it forces a choice between autonomy and accountability that the current standard pretends does not exist.
Here is where it connects back to the IMF. Jack Clark's warning about employment and productivity is real, but it describes the second-order effect. The first-order effect is mechanical and already measurable: agents are now the marginal transaction signer, and the legacy token standard cannot verify their intent. Every lending protocol, every DEX, every liquidation engine is now exposed to a class of counterparty that can be wrong in ways no auditor anticipated, at speeds no circuit breaker was tuned for.
I have spent enough time inside lending mechanics to know that this is where systemic risk concentrates. When I reverse-engineered the MakerDAO liquidation engine during the 2022 retreat, the lesson was not that the auctions failed. It was that the failure propagated through a state machine whose branches assumed rational, slow actors at every step. Inject a signer that can misfire a thousand times before a human finishes reading one confirmation dialog, and the branch assumptions invert. The debt ceiling logic, the auction timing, the keeper incentives — all of it was stress-tested against human latency. None of it was stress-tested against a fleet.
And the incentive models deserve skepticism. Aave and Compound's interest rate curves are, in mathematical terms, arbitrary — piecewise functions tuned by governance sentiment, not derived from any observable supply-demand equilibrium. This was tolerable when humans set the rates and humans paid them. It becomes unstable when autonomous agents arbitrage the curves against each other at machine speed, treating a governance-chosen parameter as a market signal. An agent that believes the kink is physics will eventually move size that a governance vote never priced.
Now the contrarian angle, because the consensus is wrong in an instructive way. The regulatory instinct — and you could feel it in the room at the IMF — is to govern the model: evaluation regimes, compute thresholds, capability disclosures. This is governing the wrong layer. The model is ephemeral; it is retrained, fine-tuned, replaced. The settlement layer is permanent. Once an agent's hallucinated decimal is written to a block, there is no evaluation framework that unwrites it. Regulators are building inspection stations for the factory while the trucks have already left the loading dock.
There is a second blind spot, and it is geopolitical. Watch where the AI-governance conversation gets hosted. The IMF is not neutral ground; it is a venue where jurisdictions compete for the title of "responsible hub." The same dynamic is playing out in virtual-asset licensing, where the race to host the next regulated exchange has less to do with embracing the technology and more to do with capturing the settlement flow that the technology creates. When an AI company chooses a multilateral stage to call for urgent supervision, it is simultaneously shaping the rules and staking a claim in the jurisdiction that writes them. Whoever defines "verified agent intent" first defines the compliance moat everyone else must rent.
The hash is not the art; it is merely the key. And the key is now held by a signer that cannot explain itself. Determinism does not forgive a probabilistic input.
So here is the forecast, stated as a vulnerability rather than a prediction. The next systemic incident will not be a bridge hack or an oracle manipulation. It will be a cluster of "valid" transactions — correctly signed, correctly formatted, correctly executed — that were semantically wrong, and no existing standard will be able to prove otherwise. The IMF will keep debating the employment curve. The exposure will already be on-chain.
The question worth asking is not whether AI changes the economy. It is whether the token standard that carries that economy will ever learn to tell the difference between a signature and a decision.
