Annex C - Concept Note on Governance of Agentic Artificial Intelligence

Annex Email: HKMA E-mail Alert of 16 October 2026 (05:00 p.m. HKT)

Back to Summaries PDF Original Document
Document Information

Title: Annex C - Concept Note on Governance of Agentic Artificial Intelligence

Type: Annex

URL: https://brdr.hkma.gov.hk/eng/doc-ldg/docId/20261015-1-EN

Email Received: 2026-10-16 18:42

Summary Created: 2026-10-16 13:09

English Summary
Quick section switch
Management Summary
  • Purpose / Background: This Concept Note outlines the HKMA’s preliminary governance considerations for "Agentic A.I."—systems capable of autonomous planning and multi-step execution. Developed from GenA.I. Sandbox findings, the note aims to guide Authorized Institutions (AIs) in managing the risks of autonomous systems while aligning with the HKMA’s "ART" (Accountability, Resilience, Trust) framework.
  • One-line conclusion: While not yet a formal supervisory requirement, AIs must begin evaluating their governance frameworks for agentic systems, focusing on intent-anchored accountability, human oversight mechanisms, and strict operating boundaries.
  • Key Changes:
  • Shift from reactive human-in-the-loop oversight to proactive "human-on-the-loop" monitoring.
  • Requirement for "intent-anchored accountability," linking agent actions directly to defined business objectives.
  • Exploration of an industry-level registry for material A.I. agents.
  • Requirement for "risk-integrated agent harnesses" to enforce operational constraints.
  • Emphasis on clear "component lineage" to manage complex interdependencies in agentic workflows.
  • Key Dates / Deadlines: N/A (Currently a consultation/discussion phase; formal supervisory requirements will follow in a forthcoming discussion paper).
  • Applicability / Impact scope: All Authorized Institutions (AIs) exploring or deploying autonomous/agentic A.I. systems that interact with internal/external systems or handle sensitive data.
  • Recommended management actions:
  • Conduct an inventory of current/planned agentic A.I. deployments and classify them by materiality.
  • Map business intent to specific agentic roles to ensure clear accountability of business owners.
  • Implement "kill switch" and emergency suspension protocols for high-autonomy agents.
  • Define clear "operating boundaries" (data/tool access) for A.I. agents to mitigate unauthorized use.
  • Prepare internal teams for upcoming HKMA engagement regarding the proposed industry-level A.I. agent registry.
Detailed Summary
  1. Document overview
  • Nature: A preliminary Concept Note on the governance of agentic A.I.
  • Purpose: To facilitate industry discussion on managing autonomous, multi-step A.I. systems.
  • Scope: Governs systems that plan, execute tasks, access tools/data, and interact with systems autonomously.
  1. Main requirements (Governance Concepts)
  • Intent-anchored Accountability: Business functions must own the "intent" of an agent. Technology functions support, but business functions remain accountable for outcomes.
  • Visibility over material agents: AIs must maintain an inventory of agents performing material functions. This includes mapping component lineage (data, models, tools) to identify incident root causes.
  • Human oversight: Transitioning to "human-on-the-loop" monitoring. AIs must have mechanisms to detect departures from business intent and execute emergency stops.
  • Operating boundaries: Agents must operate within defined constraints (tools/data access) facilitated by a "risk-integrated agent harness" to enforce internal controls.
  1. Key changes
  • Moves beyond passive oversight to requiring automated intervention points (kill switches) within agent workflows.
  • Formalizes the need for distinct "business intent" documentation as the cornerstone of risk management.
  1. Important dates & transition
  • Discussion phase: Ongoing.
  • Upcoming: The HKMA will release a discussion paper detailing potential supervisory requirements.
  1. Impact and risks
  • Operations: Increased complexity in monitoring interconnected agents; need for real-time surveillance.
  • Compliance: Potential future reporting requirement to an industry-level registry.
  • Data/IT: Risk of unauthorized data access via autonomous agents necessitating strict perimeter/harness controls.
  1. Compliance action checklist
  • [ ] Audit existing A.I. systems to identify those with "agentic" capabilities.
  • [ ] Establish a clear accountability framework (Business Owner vs. Tech Support).
  • [ ] Define "kill switch" procedures for high-autonomy agents.
  • [ ] Develop a centralized registry of material agents and their interdependencies.
  • [ ] Review current data access permissions to ensure agents only access necessary systems.
  1. Appendices/attachments summary
  • This document acts as a standalone Concept Note; there are no separate appendices. The content focuses exclusively on the core framework (Intent, Visibility, Oversight, Boundaries).
中文摘要
快速切換摘要區塊
管理層摘要
  • 目的/背景： 隨著 AI 從分析/生成內容演進為具備自主性、能執行多步驟任務的「Agentic AI（代理型 AI）」，金管局旨在透過此概念文件，引導認可機構（AIs）在享有創新紅利的同時，建立與自主性匹配的風險治理架構。
  • 一句話結論： 本文件為監管指引的初步概念，尚未成為強制要求，重點在於建立「以業務意圖為核心」的治理，並要求機構對高自主性 AI 代理進行清單化管理與人力監管。
  • 關鍵變更（治理概念）：
  1. 意圖掛鉤（Intent-anchored Accountability）： 確保 AI 行為符合業務原始目的，明確責任歸屬（業務部門而非僅技術部門）。
  2. 物質性代理可視性（Visibility）： 建立關鍵 AI 代理清單，並探討行業級註冊機制。
  3. 監管轉型： 從「人在迴路中（Human-in-the-loop）」轉向「人在迴路上（Human-on-the-loop）」的主動式監控。
  4. 操作邊界： 透過「風險整合代理束縛（Risk-integrated agent harness）」技術，將風險胃納轉化為 AI 的自動化操作約束。
  • 重要日期 / 截止日： 本文件為概念建議，監管局後續將發布討論文件（Discussion Paper）徵詢意見，目前尚無強制執行截止日。
  • 適用對象 / 影響範圍： 適用於所有使用或計劃部署自主型 AI 代理的認可機構，特別是涉及核心業務、高權限數據訪問的系統。
  • 管理層建議行動：
  1. 盤點現有 AI 代理： 建立屬於內部的「物質性 AI 代理清單」。
  2. 重塑責任架構： 指派業務部門作為 AI 代理的責任所有者，而非僅由 IT 部門擔責。
  3. 評估「殺手開關（Kill Switch）」需求： 針對高風險代理設計緊急中斷機制。
  4. 定義操作邊界： 審視並設定 AI 代理可調用的工具、數據集與系統權限。
  5. 參與監管對話： 密切關注金管局即將發布的討論文件，並透過 GenA.I. Sandbox++ 進行試驗。
詳細摘要

1) 文檔概述
本文件為金管局就「代理型 AI（Agentic AI）」治理發布的概念說明（Concept Note）。其性質屬於政策諮詢前的初步引導，旨在促進業界對話，落實「ART」框架（Accountability, Resilience, Trust），確保金融生態系統的問責、韌性與信任。

2) 主要要求（四大治理支柱）

  • 意圖掛鉤問責制： AI 代理的自主行為必須嚴格對齊業務目標。業務部門須作為問責主體，負責定義代理的授權範疇，確保 AI 產出不偏離業務初衷。
  • 物質性代理的可視性： 針對涉及關鍵功能、高自主性或存取敏感數據的 AI，機構必須掌握其「組件血統（Component lineage）」，即明確模型、數據、工具與第三方服務的依賴關係，以識別事件根因。
  • 人力監管與介入： 推行「人在迴路上」的監控模式，要求在代理執行涉及客戶利益的關鍵行動前，設置中斷點或核准機制，並配置「殺手開關（Kill Switch）」以應對偏差。
  • 操作邊界與防護： 利用技術手段建立「風險整合代理束縛（Risk-integrated agent harness）」，強制將風險胃納轉化為技術約束，防止 AI 擅自調用未經授權的系統。

3) 關鍵變更

  • 治理範式轉變： 由傳統單點式控制升級為具備「自主性認知」的控制系統。
  • 責任歸屬變化： 明確指出技術功能僅為支援角色，責任所有權歸屬於相關的「業務職能」。
  • 監管展望： 金管局正評估建立「行業級註冊登記制度（Industry-level registry）」，以便識別金融業共同依賴的 AI 模型與控制措施。

4) 重要日期與過渡安排

  • 當前階段： 屬於觀察與概念諮詢階段，非強制性監管要求。
  • 後續安排： 金管局即將發布詳細的討論文件（Discussion Paper），提出具體的監管選項，屆時將進行正式諮詢。

5) 對機構的影響與風險

  • 營運風險： 自主決策可能導致不可預期的行為鏈，增加營運崩潰或非法操作風險。
  • 資料安全： AI 代理若無嚴格邊界約束，存在濫用權限存取敏感數據的隱患。
  • 系統韌性： 多模型、多工具鏈的複雜組合，使得事故復原難度增加，需強化追蹤能力。

6) 合規動作清單（Checklist）

  • [ ] 識別並歸類機構內的「物質性（Material）」AI 代理。
  • [ ] 建立 AI 代理與業務負責人的責任映射圖（Accountability Mapping）。
  • [ ] 審查現有 AI 工作流，確認關鍵介入點（Human-on-the-loop）。
  • [ ] 測試緊急停止機制（Kill Switch）的觸發時機與程序。
  • [ ] 將「風險胃納」編程至 AI 代理的操作策略（Agent Harness）。

7) 附件/附錄摘要

  • 本文件本身為 Annex C，無其他細項附錄。文中引用的「GenA.I. Sandbox++」作為監管試驗平台，是用於驗證上述治理概念的實踐基地，並將作為未來制定具體監管要求的主要依據。
風險 / 審計心智圖 Risk & Internal-Audit Mind Map

載入心智圖中… Loading…