Hong Kong regulators test blockchain IDs that tie AI agents' payments to verified customers

Know-your-customer checks establish who a customer is. When that customer's AI agent asks for a transfer, the firm still has to confirm separately that the customer gave the agent authority to send money, and how far it may go ^[6]. The Sedaily report describes this as the gap the new IDs are meant to close ^[6]. HKT Payment's version is built to prove on whose behalf an agent is acting ^[5], and the category's broader aim is to keep agents inside approved limits ^[3]. A firm holding such IDs could refuse requests from agents whose identity or authority does not verify ^[11].

The rest of the field is organised differently. Of the four efforts the report describes, only HKT Payment's runs inside a regulator's program ^[15], and it pairs one payments firm with one blockchain supplier ^[4]. Google is involved in two of the other three, through the ERC-8004 proposal and the Mysten Labs arbiter ^[16]. Hashed chief executive Kim Seo-joon wrote in February that agents from different companies need a neutral basis of trust that no single company controls ^[7]. Blockchain, he argued, lets several companies check an agent's identity, reputation and task data within one shared framework ^[7]. In an April report, a16z crypto said there was still no widely used method to prove on whose behalf an agent acts or what authority it holds ^[14].

Three outcomes fit the evidence. Hong Kong's regulators ^[1] could turn the binding of agent to customer into guidance, so that supervised firms have to show a mandate check for every agent payment. They could accept an open standard such as ERC-8004 ^[9], leaving no single vendor in control of the identity layer. Or the sandbox could close as a test. Spending limits would then stay a feature each bank designs and prices for itself.

We think the first is the most likely in Hong Kong. A regulator is running the test ^[1], and the gap it targets sits with the financial firm, which carries the authority check today ^[6]. The counter-case is that a sandbox is an experiment. The report does not give the pilot's cost or timeline, or say whether spending caps are part of what the regulator is testing. If the program ends without a stated expectation on agent mandates, or if banks adopt open spending-limit standards before the regulator says anything, this view is wrong, and agent limits remain a product decision.