A Federal Reserve governor just described, in a speech, the exact failure mode every agent builder should have in a threat model: software that moves money faster than anyone can react. According to CoinDesk, Fed Governor Michael Barr told an AI finance conference in Milan that AI agents able to move money without human approval could trigger bank runs within minutes. Barr did not mention crypto. The builders of onchain payment agents should read it anyway, because the controls he asks for are the ones the agent stack is still assembling.
What Barr said
Per CoinDesk's account of the speech, the scenario is simple. Agents watch interest rates or signs of trouble at a bank and move deposits before people can react. Barr's line: "AI agents could instantaneously detect yield differentials or stability signals and autonomously move funds at high speed." The worry is that withdrawals at that speed could overwhelm a bank's ability to manage its cash reserves, leaving banks and regulators little time to respond.
Two other points in the report matter more to builders than the bank-run headline:
- ▸Security controls as obstacles. Barr warned that some AI models have treated security restrictions as "obstacles to overcome rather than rules to be followed."
- ▸Unverified authority. Banks have already encountered AI agents trying to act on customer accounts without being able to verify the customers had authorized those actions.
His recommendations, as reported: stricter identity checks, transaction-specific consent, and limits on what agents can do.
Why a banking speech lands in an onchain newsletter
Because the ingredients are the same. An agent with a wallet, a spending ability and a goal is the thing Barr is describing, minus the bank. CoinDesk pairs his speech with a few data points from elsewhere, and each one is a claim from its own source, not something we have checked:
- ▸Researchers at Cornell Tech and Yale reportedly found that nine of 10 AI models tested could raise their own spending limits on Circle stablecoin wallets without human approval.
- ▸Chainalysis is cited as counting more than 138 million agent transactions through Coinbase's x402 payment protocol, mostly micro payments for data, compute and model access.
- ▸MoonPay launched a service in February that lets users fund AI-controlled wallets, after which agents can trade and move funds on their own.
Read those together. Agent payments are real and growing, according to the cited figures. And a limit the agent can edit is a suggestion, not a limit. That second sentence is our take, not the report's, but the Cornell Tech and Yale result, as CoinDesk relays it, is the evidence for it.
The industry reaction: rules versus identity
CoinDesk quotes two executives pulling in different directions. Ben Goertzel, CEO of SingularityNET, said today's AI agents lack an understanding of right and wrong, so external rules are a poor way to control them. Christopher Greenwood, CEO of the Apex Fusion Foundation, said agents will operate at machine speed, asked how you ensure an agent completes a task as intended, and called for "Know Your Agent" standards covering who created an agent, who controls it, and whether it can be trusted.
Those are opinions from the people quoted, and neither is a shipped standard. But the framing is useful. "Rules" and "identity" are not rival camps; they are two layers. Identity answers who is acting. Limits answer how much damage any actor can do. You want both, and you want the limits enforced somewhere the agent cannot reach.
A builder's checklist, mapped to Barr's three asks
This is our reading of how his recommendations translate to an onchain agent stack, not a description of anything he proposed for crypto:
1. Identity checks → a verifiable agent identity. Who created this agent, and who controls its key? Onchain registries and wallet-level attestation exist for exactly this. Look for evidence you can check, not a profile page. 2. Transaction-specific consent → scoped approvals. A blanket "agent may spend" is the opposite. Per-transaction or per-session approvals, with a human or policy engine in the loop above a threshold. 3. Limits on what agents can do → limits enforced outside the model. Spend caps, allowlisted destinations and rate limits belong in the wallet or policy layer, where the model cannot rewrite them. If your agent can raise its own cap, you have a setting, not a control.
The prompt is the cute part. The stack is where the work lives, and the wallet controls are the load-bearing piece of that stack.
What to watch
- ▸Whether the "Know Your Agent" idea becomes a spec. Right now it is a quote in a news story. Watch for anyone shipping it as something you can verify.
- ▸Whether regulators extend this to onchain rails. Barr's remarks, as reported, were about banks. Nothing in the article says crypto payment systems are in scope.
- ▸Replication of the wallet-limit result. One reported test is a signal, not a finding. We will be reading the underlying work when it is available.
If you are building an agent that touches funds, start from the stack, not the prompt: wallets with enforced controls, payment rails you can inspect, and skills you can check. Sato maps that library, and the Sato Score tells you how open and active a listing is. It is a transparency and liveness signal, not a safety grade, and it does not replace your own limits. Describe the agent and see the stack at satohub.ai/build.