Hong Kong creates first banking rules for AI agent identities

Closing the autonomy gap

Hong Kong's central bank recently debuted Project SAGE, a first-of-its-kind governance initiative for autonomous AI agents in banking. The project addresses a structural gap in current financial regulations where software can perform transactions, handle legal agreements, and direct other AI programs without a person approving every single action.

Existing rules like Know Your Customer (KYC) were designed for people or registered companies, not software. Because AI agents can be copied, modified, or deleted instantly, they lack a standard way to prove who deployed them or what they are permitted to do. Project SAGE aims to create a formal identity credential for these digital actors.

The Hong Kong Monetary Authority (HKMA) developed this framework to manage the risks of multi-agent orchestration. In these setups, one AI agent might instruct a sequence of sub-agents to perform complex tasks across different currencies and time zones. This creates a series of hand-offs that current oversight systems were not built to track.

How 'Know Your Agent' works

The core of the initiative is the 'Know Your Agent' (KYA) pillar, which focuses on designing identity credentials for AI. These credentials must include cryptographically signed certificates that attest to an agent's authorized scope and specific capabilities. This allows counterparties to verify exactly what an agent is allowed to do before a transaction occurs.

Beyond identity, the framework requires behavioral audit logs to reconstruct multi-step decision chains. This is necessary because agentic AI is architecturally different from generative AI; it plans a sequence of actions, executes them, and replans based on the outcome. These logs ensure that if an error occurs, the sequence of autonomous decisions can be audited by regulators.

The HKMA is also looking at interoperable standards for cross-border transactions. If an agentic payment moves from Hong Kong to the Chinese mainland or a digital asset layer, shared technical protocols are required. These protocols ensure that liability can be traced even when the transaction involves multiple jurisdictions and different types of settlement layers.

Preventing agentic bank runs

Regulators are concerned that unregulated AI agents could create a new type of systemic risk known as an 'agentic bank run.' Economists warn that if agents are deployed to optimize household finances, they could simultaneously move deposits into higher-yield products. This would eliminate the behavioral inertia that historically kept retail funds in place even when better yields existed elsewhere.

Unlike traditional bank runs driven by panic, an agentic bank run would be driven by optimization. Because agents remove the friction of manual transfers, they could move large sums of money instantly. The Cloud Security Alliance recently classified this as a systemic financial stability risk rather than a simple operational risk for individual banks.

To counter this, Project SAGE includes a proposal for a centralized registry of 'material' AI agents. The HKMA is also considering 'kill-switch' powers, which would allow the regulator to suspend specific agent IDs at the network level. This is architecturally similar to the circuit breakers used to halt disorderly markets during periods of extreme volatility.

Evidence from the sandbox

Project SAGE was not built in a vacuum but is grounded in data from the HKMA's GenAI Sandbox++ program. This program chose 36 specific use cases involving 30 different financial firms and 27 technology providers to test AI in controlled settings. The data showed that 60 percent of the proposals in the previous round were already focused on agentic AI rather than generative tools.

The sandbox pilots provided direct evidence of the governance gap as banks began building autonomous capabilities at pace. The HKMA observed that human-only oversight is practically unworkable at the transaction speeds and complexity levels of multi-agent architectures. This led to the exploration of 'AI versus AI' risk management, where one system monitors the actions of another.

The initiative also builds on the HKMA's ART Framework, which stands for Accountability, Resilience, and Trust. This framework requires clear human ownership throughout the AI lifecycle and ensures systems fail safely. Project SAGE is the practical implementation of these principles, moving from high-level values to specific engineering and supervisory designs.

The global regulatory race

Hong Kong's move puts it in direct competition with Singapore for leadership in AI governance. While Singapore released a broad, sector-agnostic principles framework in early 2026, Hong Kong is focusing on 'engineering plumbing.' The HKMA is asking specific questions about credential structures and transaction thresholds rather than just drafting general values.

The project is part of a broader strategy involving data, payments, and tokenization. Alongside the AI pillar, the HKMA is advancing tokenized settlement pilots and post-quantum cryptography toolkits. These tools ensure that the AI agent infrastructure remains secure against future quantum computing threats that could decrypt captured communications.

As the private sector moves forward, startups are already raising millions to build KYA-grade credentialing systems. By providing a formal mandate, the HKMA is setting a precedent for how banking regulators will manage the shift toward autonomous finance. This framework provides a blueprint for how institutions can pair innovation with effective human oversight.