This is a static, simulated illustration. The agent, the purchase and the messages are invented, and nothing is really bought or cancelled. The explainer page is linked from the top.
Agent Escrow is an illustration of a capability a declared agent would get for declaring, namely, the ability for the agent owner to revoke an action. The underlying idea is that rather than preventing agents from making mistakes or micromanaging agents, we build in mechanisms for reversing actions easily. Using the example of a purchase, Agent Escrow functions similarly to a worktree: real actions take place in a temporary branched workspace. Actions in this workspace can be undone and fully deleted without consequence (and in the example within a certain period of time) while actions can also be accepted and take effect.
Agents can now act for people, including spending their money, and when an agent does something the person didn't expect, the only recourse is a refund request or chargeback after the order is already final.
People who let agents act for them, and the merchants who take those actions, have this problem. Agents can complete tasks end to end without ongoing oversight, and protocols such as AP2 and TAP help constrain what they can do and spend. But as agents become more capable and users are exposed to opt-out features, accidental purchases and actions will happen.
Nothing in the transaction tells the merchant whether a person or an agent made it, so the merchant has no reason to treat agent purchases differently, and the agent has no reason to say what it is. It matters because the only recourse today is a refund request or chargeback after the order is final.
Rather than trying to prevent all unexpected actions, which is impossible, or micromanaging agents, Agent Escrow builds in a way to reverse actions easily. It works like a worktree: real actions happen in a temporary branched workspace that can be thrown away until someone accepts them. The person is told right away what was done, what it cost and the deadline. A reply of ctrl-z revokes the action, and any other reply confirms it early.
The window is what a declared agent earns by declaring: it tells the merchant that software is acting for a person who can be reached, which gives the merchant a reason to hold the order open longer. Reversal fits things that haven't been used or delivered yet, such as tickets, travel, goods that ship and digital artifacts. It does not fit actions that cannot be undone, such as a sent message or a service already consumed.
A person who can revoke an agent's real-world action within a window is more willing to let the agent act, and a site that sees the agent declared has reason to grant the longer window.
If that holds, declaring becomes the way an agent earns the revocation window, and agents can complete real tasks end to end without the person having to prevent every mistake in advance.