This year, software started to carry money on its own.
In February Coinbase launched wallets built for AI agents, with programmable spending limits and session controls.1 In July x402, an open protocol that lets programs pay for what they use over the web, got its own foundation at the Linux Foundation, with forty member organizations.2 Two days later Ledger released an open-source toolkit that lets agents prepare transactions while a person approves the sensitive ones on a hardware device.3 This week Mastercard opened developer access to a pilot for agent payments built on x402, describing it as payment without “per-transaction human sign-off.”4
The rails settle how an agent pays. They don’t settle what the person sees before it does, or what happens when the agent is wrong. That part is design, and most of it hasn’t been done.
The decision has moved
For fifteen years I’ve worked on screens where a person hands power to a machine: a flight booking, a trading terminal, a wallet’s swap. In all of them the person pressed the button. The interface had one job: make a single decision legible at the moment it was made.
Agents move the decision away from the moment. You decide on Monday what an agent may do, and it acts on Thursday while you sleep. Most of the design work moves to two places today’s products barely have: the rule you set before, and the record you read after.
Five things a person needs to see
When an agent does ask, the request has to answer five questions, in the same order every time:
- What it wants, in plain words. Not a contract call. “Pay $12.40 to flights-api.dev for 400 fare lookups.”
- Why. Tied to the task the person gave. An agent that can’t say why shouldn’t be allowed to ask.
- Which rule it falls under. A limit the person wrote and can read, and how much of it is already used.
- What could go wrong. The worst case in money, not adjectives.
- How to undo it. A cancel window, a refund route, or an honest “you can’t.”
You can try a working sketch of this in Principles that run.
Approval fatigue is the real enemy
If every action needs a tap, people approve everything. Wallets have lived with a version of this for years: people grant token approvals they don’t understand because the approval sits inside a bigger flow. At Trust Wallet we pulled approval out of the swap so it could be read on its own.
Agents multiply the problem. The answer can’t be more prompts. It has to be better rules, written once in the person’s own terms, with interruptions saved for what the rules don’t cover.
Refusal is a feature
The most useful thing an agent can do is say no on its own. When a request breaks a rule, a good agent stops before it asks, and says which rule, so the person can decide whether the rule or the request was wrong. An agent that can refuse is one you can trust with more.
In DEED, the agent treasury I built in 2026, every refusal is kept with the rule it protected. Those records are how the rules get better.
Some moments need a person, not a session
Raising a limit, adding a recipient, moving everything: these should need a signature only the person can give, on a device in their hand, and a delay a hijacked session can’t skip. Hardware wallets already work this way. Ledger has put a proof-of-human attestation on its roadmap for agents, and named a Chief Human Agency Officer.5 The idea is about to go mainstream.
The receipt is the interface
After the fact, the record is the product. A receipt should say what was proposed, under which rule, who approved it, what happened, what it cost and how to dispute it. If people can’t read it, they can’t correct the agent, and the rules never get better.
What I’m doing about it
I’m publishing these as Agency Patterns, an open library: each pattern with the problem, the pattern, where something like it already exists, and a test for whether it works. Version 0 has six. I’m building DEED to test them with real agents. And I work with teams building wallets, payments and agents who want this done properly before their users find out the hard way.
If you’re one of them, write to me.
Sources
- Payments Journal: Coinbase unveils Agentic Wallets, February 2026
- x402: Linux Foundation announces the operational launch of the x402 Foundation, 14 July 2026
- The Block: Ledger unveils a hardware-backed agent stack, 16 July 2026
- Tech Times: Mastercard opens developer APIs for AI agent payments built on x402, October 2026
- Ledger: 2026 AI security roadmap; eCommerceNews: Ledger appoints its first Chief Human Agency Officer, April 2026