The Short Answer
Know Your Agent, in the HKMA's usage, means verifying an AI agent's identity and its delegated authority before it acts in finance, and it is still being explored, not mandated. A bank will have to answer for its agents from its own records, which is why ibl.ai deploys the agent stack where you own all the code and the data.
On 2 November 2026 the Hong Kong Monetary Authority (HKMA) launched Project SAGE, short for "Securing Agentic Finance with Governance and Enablement". Agent identity is one of its four workstreams.
The question is real. The popular version of it, "a regulator asked banks how to verify an agent, and nobody has an answer", is not quite what happened. The sections below separate the two.
What did the HKMA actually announce about Know Your Agent?
The HKMA's press release of 2 November 2026 lists four focus areas for Project SAGE. The second is agent identity.
In its words, the HKMA will "explore approaches to verifying A.I. agents' identities and delegated authority, including Know Your Agent (KYA)" practices and authentication, so that agents can take part in financial activity in a trusted way.
| Project SAGE workstream | What the HKMA says it covers |
|---|---|
| 1. Market development and use cases | B2C agentic commerce and payments, B2B procurement and treasury, machine-to-machine usage-based payments |
| 2. Agent identity | Verifying an agent's identity and delegated authority, including KYA practices and authentication |
| 3. Interoperability standards | Standards and protocols for agents to interact across the ecosystem |
| 4. Legal and regulatory frameworks | Governance arrangements, liability and consumer protection |
The work is supported by a new Agentic Finance Expert Group. Its published member list has 22 entries: 11 financial institutions, 7 technology providers, 3 professional services firms and 1 academic.
Among them are HSBC, Standard Chartered Bank (Hong Kong), Bank of China (Hong Kong), Visa Hong Kong, Mastercard, Ant International, Adyen, AWS, Google Cloud, Tencent Cloud, Clifford Chance and PwC.
Caixin reported that HKMA Chief Executive Eddie Yue announced the project at Hong Kong FinTech Week, and that the group is expected to produce preliminary results, including demonstrations and white papers, within 9 to 12 months.
Is Hong Kong requiring Know Your Agent checks for banks yet?
No. Project SAGE is a programme to explore Know Your Agent, run with industry, and nothing in the HKMA's release imposes a KYA requirement on any bank today.
Some coverage has described Hong Kong as already requiring identity credentials for AI agents. The HKMA's own documents describe exploration and consultation, so we read those headlines as running ahead of the source.
The supervisory direction is clearer in an earlier step. On 15 October 2026 the HKMA announced its "ART" framework and an upcoming discussion paper on a supervisory framework for agentic AI governance.
That paper, the HKMA said, will explore three things: intent-anchored accountability, a registry of material AI agents, and effective intervention mechanisms. It had not been published when this post went out.
Why do banks need Know Your Agent now, and not in five years?
Because AI agents are already the majority of use cases in the regulator's sandbox. In the second cohort of the HKMA's GenAI Sandbox, more than 60% of use cases involved AI agents, and half of those were multi-agent applications.
That figure comes from remarks by Carmen Chu, the HKMA's Executive Director (Banking Supervision), at the Finance + A.I. Forum on 16 October 2026.
In the same speech she said that for agentic AI "governance must be anchored to intent", and that intervention measures may include "kill-switch" mechanisms to suspend agents before an issue escalates, where appropriate.
She also said the HKMA will explore whether a centralised supervisory registry could support this across the banking sector. A registry presupposes that each bank can say, precisely, which agents it runs.
Is there really no identity infrastructure for AI agents today?
There is some, and it is worth being exact about it. Agent identity is not starting from zero; what is missing is a supervisory standard that tells a bank which proof is enough.
Mastercard announced Agent Pay on 29 April 2025 and said the programme "will require trusted AI agents to be registered and verified". It issues Agentic Tokens built on its existing tokenisation.
Inside the enterprise, Microsoft Entra Agent ID gives agents their own identities to authenticate, authorise, govern and protect. Extending Entra's security features to agents requires Microsoft Agent 365, licensed per user.
| KYA question a supervisor will ask | Where an answer exists today | What the bank still has to hold itself |
|---|---|---|
| Which agent is this? | Card-network agent registration (Agent Pay, announced April 2025); directory identities such as Entra Agent ID | An inventory mapping each identity to a deployed agent and version |
| Who delegated authority to it? | Consumer authentication at payment time; OAuth-style consent | The mandate record: which customer or employee, for what, until when |
| What was it trying to do? | The HKMA's "intent-anchored accountability", planned for a forthcoming discussion paper | The prompt, plan and tool calls behind each action, retained and reviewable |
| Who could have stopped it? | "Kill-switch" mechanisms, which the HKMA says intervention measures may include where appropriate | An intervention path the bank itself controls, not one it requests |
The left column is the network's or the platform's answer. The right column is the bank's, and no third-party identity token supplies it.
Who should hold the records that prove an AI agent's identity?
The bank should, because the bank is the one the supervisor asks. A KYA answer is assembled from agent definitions, permissions, mandates, logs and intervention controls, and those sit wherever the agent runs.
If the agent runtime is a managed service, much of that evidence lives in the vendor's system, under the vendor's retention period, export format and contract. That can work. It does mean the bank's answer to its supervisor depends on a third party's records.
ibl.ai is built for the other arrangement. You own all the code and the data: the agent definitions, the permission model and the audit trail run in your own cloud, on-premise or air-gapped, under source code you hold.
It is model-agnostic, so a model swap does not change the record.
It is usage-based, with no per-seat pricing, so an inventory of hundreds of agents does not turn into a headcount bill. The pattern for regulated institutions is on our financial services solutions page and the Agentic OS page.
What should a bank prepare for Know Your Agent before the HKMA paper lands?
Start with what the HKMA has already named. Its October and November documents point at the same four records, and each can be built before any rule is written.
- An agent inventory. Every agent in production, its owner, version and the systems it can touch. The proposed "registry of material AI agents" will assume one exists.
- Delegation records. For each action, whose authority the agent used and its limits. Permissions scoped at the row level are one way to keep that narrow, as AWS's TOLAP release showed.
- Intent trails. The request, plan and tool calls behind each action, kept long enough to review, since the HKMA wants governance "anchored to intent".
- A stop control the bank operates. The HKMA said intervention measures may include kill switches where appropriate, so test that you can suspend one agent without taking down the rest.
None of these is new technology. They are the agent-era version of KYC files, and the identity layer itself has governance and ownership questions of its own.
ibl.ai is family-owned and operated from New York, NY. To see a governed agent stack running inside your own infrastructure, book a 30-minute demo or talk to the ibl.ai team.
Sources: Project SAGE and its four workstreams from the HKMA press release of 2 November 2026 and its member annex; the timeline from Caixin; the ART framework and planned discussion paper from the HKMA press release of 15 October 2026; sandbox figures, intent and kill-switch remarks from Carmen Chu's speech of 16 October 2026; Agent Pay from Mastercard; Entra Agent ID from Microsoft Learn.