Suit Up · Asia's institutional desk on tokenisation

The current regulatory position on AI agents holding or moving tokenised money is, almost everywhere, silence. No major jurisdiction has published primary regulation that addresses agent-mediated transfers as a distinct category. The treatment runs by inference from existing principal-and-agent doctrines, licensing perimeters that attach to the principal, and AML and KYC frameworks drafted with human and corporate holders in mind. The practical effect is that agentic flows are operationally workable today on almost every regulated rail, with the perimeter mostly silent rather than hostile, and the legal hook resting on traditional agency law that has not yet been tested in any significant agentic dispute. This page maps where the published positioning sits, where the gaps are, and what we think gets answered first (Foundations Chapter IX, Part 4 is the longer-form treatment).

The default perimeter

Every major payment-stablecoin regime that settled into operative form across 2024 and 2025 (the US GENIUS Act, MiCA, the Hong Kong Stablecoins Ordinance, the Japanese PSA amendments, the MAS SCS framework) places regulatory weight on the issuer rather than the holder. None impose holder-level KYC at the chain level. None contemplate AI agents as a distinct holder category. Chain-level transferability runs to any address not on the issuer's blocklist, regardless of whether the address is controlled by a human, a corporate, or an agent acting under delegated authority. The same posture extends to tokenised deposits: the bank's KYC runs on the principal, the chain enforces the bank's whitelist, and whether the principal's authorised agent initiates a transfer is operationally a matter for the bank's internal policy rather than for the regulatory text.

KYC and AML obligations attach to the principal in every regime we have reviewed. A bank onboarding a corporate principal runs KYC on the corporate, documents the authority granted to the agent, and accepts the resulting transactions as transactions of the corporate. The unanswered question is whether sanctions-screening obligations attach to the agent's reasoning process or only to the transaction at the chain level. Our working read is the latter, but no regulator has published a position.

The legal frame is principal-and-agent. Common-law doctrines and their civil-law equivalents in Japan, South Korea, Hong Kong's hybrid system, and Singapore treat an agent as binding the principal when acting within the authority granted, regardless of whether the agent is a natural person. Apparent authority protects a counterparty who reasonably trusts an agent's outward-facing claim even when the principal's instructions are narrower. The doctrines have answers in principle for malfunction, hallucination, and acts outside authority. The case law on LLM-driven agents is thin to non-existent.

Property characterisation

Whether tokenised money held by an AI agent is treated as agent property, principal property, or a new category is the structural question that has not been resolved. Across every framework reviewed the operative answer is principal property: the agent is a tool of the principal, wallet credentials sit under the principal's control or under a custody arrangement the principal authorised, and legal title runs to the principal. The agent has no independent property right. An agent acting outside authority through prompt injection or hallucination is handled by apparent-authority and ostensible-scope doctrines; an agent receiving a tokenised payment in error is handled by existing law on mistaken payments and unjust enrichment. None of this has been tested in a published agentic-specific dispute.

The product-team takeaway is that agent-compatible product design treats the agent as a transparent tool of the principal. The licensing weight runs to the principal. There is no need to give the agent its own KYC, account, or legal identity. Memos that propose to "license the agent" or "give the agent KYC" are misreading the architecture.

What the regulators have said

The published positioning is thin. The MAS under Sopnendu Mohanty's tenure has been the most agent-curious of the major APAC regulators, with public commentary across 2024 and 2025 acknowledging that the convergence of AI agents and programmable money is the next architectural surface and that Project Orchid purpose-bound money positions Singapore well to address it. No specific MAS guidance has been published as of April 2026. We read the Project Orchid PBM construct as the most agent-aligned existing piece of MAS infrastructure: PBM conditions a token's transferability on specified conditions enforced at the contract level, which is exactly the structure an agent acting under conditional authority needs.

The HKMA has not published explicit positioning on agentic flows in EnsembleTX or the broader Project Ensemble architecture, but industry coverage through late 2025 and into 2026 has carried unverified reports of HKMA exploring agent-pilot extensions to the existing tokenised-deposit work. We treat these as not yet load-bearing and flag the absence of a primary-source HKMA statement. EnsembleTX's structural posture (bank holds principal-side KYC, chain enforces the bank's whitelist) accommodates agentic flows in principle without new regulatory text.

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 BIS Innovation Hub under Tommaso Mancini-Griffoli's leadership starting March 2026 is the most likely venue for a multi-jurisdictional pilot on agent-mediated wholesale settlement if any materialises.

The US picture is more fragmented. The CFTC algorithmic-trading framework (the never-finalised Reg AT proposal) is the closest existing federal template, with provisions on automated-trading firm registration, source-code preservation, and compliance-officer designation. The framework was withdrawn before finalisation and was scoped to derivatives trading rather than spot tokenised flows, but the design choice (registration and compliance attaching to the operator of the automated system rather than to the system itself) translates cleanly to the agent context. The CFTC has not extended the design to digital-asset commodity exchanges under CLARITY. The SEC and the OCC have likewise been silent.

Where the gaps are

Three gaps we treat as operationally consequential. First, sanctions and AML reasoning-process responsibility. Whether an agent that runs a sanctions-screening prompt internally before signing a transaction is treated as having made a sanctions-relevant determination, or whether the obligation attaches only to principal onboarding and the chain-level transaction, is undisclosed across every major jurisdiction. The first enforcement action will set the pattern.

Second, agent-mediated trigger of significant-instrument thresholds. MiCA's significance thresholds for ARTs and EMTs, the HK Ordinance's analogous mechanics, and the Japanese PSA's holder caps were drafted against human and corporate holder counts. High-frequency agentic flows could trip thresholds in ways the drafters did not contemplate. Whether regulators interpret thresholds against the operative legal holder or against the chain-level apparent counterparty count is undisclosed.

Third, dispute resolution when the agent acts outside authority. The doctrines have answers in principle but have not been tested at scale on LLM-driven failures (hallucination, prompt injection, scope creep in pursuit of a higher-order goal). The first major case will likely shape underwriting, insurance products, and product-side authority-management design choices for several years.

What we think gets answered first

MAS is the most likely to publish an explicit accommodation framework for agent-mediated flows on a 12 to 24-month timeline. The Project Orchid PBM construct is the obvious vehicle, and the integrated-regulator structure makes single-published guidance covering banking, payments, and capital markets more achievable than the multi-agency US or Hong Kong configurations. The HKMA is the most likely to ship a regulator-coordinated agent pilot inside the EnsembleTX architecture on a similar timeline. The first major dispute over agent-driven tokenised-money loss is more likely to surface in the US, where the chain-native pattern has the most production volume and the legal infrastructure for novel financial-technology disputes is deepest.

We do not expect agent-specific licensing to emerge in any major jurisdiction in the near term. The path of least resistance is principal-side licensing extensions on what authority a principal can grant, what disclosures are required, and what audit trails must be kept. Agent-specific licensing would require building a new regulatory category from scratch; principal-side extensions plug into infrastructure that already exists.

Open questions

Which regulator publishes the first explicit text addressing agent-mediated transfers in regulated tokenised money. MAS, the BoK under DABA Phase 2, the EBA under MiCA implementation, and the BIS through the Innovation Hub are the candidates.
Whether the GENIUS Act perimeter produces any FDIC, OCC, or Federal Reserve guidance on bank-issued stablecoin or tokenised-deposit acceptance of agent-initiated transfers.
The substance of any HKMA pilot on agent-mediated EnsembleTX flows. A primary-source HKMA statement would resolve.
How the FSA Japan AI Discussion Paper Version 1.1 treats AI agents holding or moving regulated digital money in Japan (flagged in outputs/_questions.md).
Whether the first major agentic-commerce dispute surfaces in a regulated tokenised-deposit context or in a stablecoin context, and which jurisdiction's courts get the first case.

[Suit Up]

Regulatory perimeter for agents