The regulatory question is not whether agentic commerce is permitted. The question is whether the regimes that govern each tokenised instrument accommodate, ignore, or implicitly block agent-mediated flows. Most regimes ignore them at the text level. A product team building today is operating in the silence, not in the prohibitions.
Regulatory accommodation
Most regimes ignore AI agents at the text level. The hong kong stablecoins ordinance, the MiCA text, the japan psa stablecoin routes, and the singapore mas scs framework are all silent on agents. The HK Ordinance and the MAS Digital Token Service Provider regime have intermediary custody provisions that probably extend to agentic custody where the agent custodies tokens on behalf of a human principal, but the question has not been litigated and no regulator has published guidance.
The GENIUS Act payment stablecoin perimeter is consistent with agentic use because it does not impose holder-level KYC at the chain level; the perimeter is on the issuer. MiCA's per-instrument significance thresholds for asset-referenced tokens (ARTs) and e-money tokens (EMTs), plus caps on use as a means of exchange for non-EU-currency stablecoins, are the pressure points if agentic flows scale. Agentic flows that look like person-to-person flows but run across many counterparties at high frequency could trigger thresholds in ways the drafters did not contemplate.
The MAS Project Orchid work has explored purpose-bound money (PBM) structures that map cleanly to agentic conditional payments. PBM conditions a token's transferability on specified conditions (recipient identity, use case, time window). The structure is exactly what an agent acting under conditional authority needs: the principal grants a PBM the agent can spend only on specified categories, with conditions enforced at the contract level rather than relying on the agent's reasoning. Singapore is, on the available evidence, the most agent-curious major APAC regulator, with its Project Orchid and DTSP work positioning MAS to be first to address agentic flows specifically if any major regulator does.
The BIS staked out the macro frame in its Annual Economic Report 2024 chapter on AI and tokenisation, arguing that the convergence of agents and programmable rails will be operative before regulators write text addressing it. The chapter sits as the closest thing to a regulator-side acknowledgement that agentic flows are coming faster than text can be drafted.
Product-team takeaway
Agent-to-agent settlement on tokenised rails is technically achievable today on multiple stacks. Stablecoin transfers between two agent-controlled wallets work end-to-end on the Coinbase stack and on the Solana, Base, and other major USDC-supporting chains. Tokenised deposit transfers between agent-authorised accounts at the same bank work on Kinexys today; cross-bank flows in the substitution pattern work on Partior. Tokenised MMF subscription and redemption by an agent acting for an eligible principal works on BUIDL and FOBXX today, subject to onboarding requirements.
For a tokenisation product team, the right question now is whether the product is agent-compatible without yet being agent-specific. Agent-compatible means chain-level transferability that does not require human-in-the-loop for every transfer, programmable conditions at the smart-contract level, machine-readable APIs for issuance and redemption, and a documented authority model for delegated transfers. Agent-specific means betting on a particular agent stack. The first is the conservative bet, accommodates whatever agent ecosystem materialises, and is achievable today on most permissioned chains and all the major public chains. The second is a venture-grade bet on a particular winner.
Why it matters now
Agentic commerce and tokenisation are converging on a timeline that no major regulator has yet caught up with. The instruments are ready. The chains are ready. The legal framework is workable, if untested in agentic-specific disputes. The architectural primitives have emerged. What is missing is regulatory text that addresses agentic flows specifically, and the absence creates a window in which product teams can ship without bumping into hard text-level prohibitions, but also without the comfort of explicit accommodation.
Two structural facts make the timing pointed. The major payment stablecoin perimeters (GENIUS, MiCA Title IV, the HK Ordinance, the Japanese PSA, the MAS framework) all settled into operative form across 2024 and 2025 with chain-level transferability preserved, which means the rails are ready for any holder type the regimes do not explicitly exclude. The Project Ensemble tiered ledger architecture and the equivalent work in Singapore, Japan, and south korea has put tokenised deposits into production, which means client-facing programmable money is no longer a research project. An agent acting for a corporate treasury today can route incoming USDC into BUIDL, hold a tokenised deposit at a participating bank, and pay for an x402-monetised API call, all without leaving the regulated stack.
The wedge sits at the intersection of these two facts. It is not a thesis about which agent stack wins. It is a thesis that the regulated tokenised stack has accidentally become agent-ready, and the regulatory perimeter is silent on the agent question in ways that will become operative as flows scale.
Common confusions
AI agents are not legal persons. They cannot be the legal subject of a transaction. The transaction is legally the principal's, with the agent as a tool. Memos that propose to "give the agent KYC" or "license the agent" usually misread the legal architecture; KYC and licensing attach to the principal.
Agent-compatible is not agent-specific. A product that works for human and corporate holders, with chain-level transferability and programmable conditions, is already agent-compatible in the relevant sense. There is no need to bet on a specific agent stack to be ready for agentic flows.
Agentic flows are not necessarily retail flows. The first wave of agentic commerce on tokenised rails is largely corporate treasury, machine-to-machine API settlement, and B2B conditional payments. The retail agentic case (an LLM acting as a personal financial agent for a household) is later, more demanding on consumer protection grounds, and not where the volume will land first.
Verifiable credentials are not blockchain-specific. The W3C DID and VC standards are chain-agnostic, and the architecture works equally well with off-chain identity providers, traditional certificate authorities, and on-chain issuance.
The agent does not need a wallet to act on behalf of the principal. In the Stripe Agent Toolkit, Visa Intelligent Commerce, and Mastercard Agent Pay patterns, the agent operates against the principal's existing payment relationship without holding any tokens itself. Both patterns are agentic; the chain-native pattern is one of several.
A blocked stablecoin address does not block the agent. Issuer blocklists block specific addresses, not the agent's reasoning process. An agent that loses access to one address through a blocklist event will, in most operational designs, route to a different address controlled by the same principal, subject to that address not also being on the list.