Docs

Appearance

Singapore IMDA Model AI Governance Framework for Agentic AI (v1.5) ​

The MGF for Agentic AI ​

The Model AI Governance Framework for Agentic AI (MGF for Agentic AI) is voluntary guidance from the Infocomm Media Development Authority (IMDA) of Singapore. The MGF for Agentic AI is not law. It sets no obligations in law. The IMDA press release of 20 May 2026 tells organizations that the MGF for Agentic AI "provides guidance to organisations on how to deploy agents responsibly, recommending technical and non-technical measures to mitigate risks" (¶1).

The MGF for Agentic AI is for organizations that want to deploy agentic AI. These organizations can make AI agents in the organization or use "third-party agentic solutions" (Executive Summary).

The pages about the MGF for Agentic AI use Version 1.5. The first page of Version 1.5 gives the dates "Published 20 May 2026" and "Updated 5 June 2026". In Annex B, IMDA tells organizations: "This is a living document, and we invite suggestions on how the framework can be updated or refined." Annex B also has a "Call for case studies".

The Executive Summary tells organizations that "principles for trusted AI" "continue to apply", for example transparency, accountability, and "fairness". IMDA adds that these "principles" "need to be translated in practice for agents" (Executive Summary, p. 3).

The MGF for Agentic AI "builds on the responsible AI practices for organisations set out in MGF (2020)" (§2). Footnote 16 identifies MGF (2020) as the "Model AI Governance Framework (2nd Ed)". The MGF for Agentic AI also refers to other documents, for example the IMDA case study on OpenClaw (§2) and the "Starter Kit for Testing of LLM-based Applications for Safety and Reliability" (§2.3.2).

The four dimensions of the framework ​

The MGF for Agentic AI gives organizations four dimensions (§2). Each dimension has a page:

Dimension and pageInformation on the page
1. "Assess and bound the risks upfront" (§2.1)Risk assessment of agents, impact and likelihood, limits on agents, agent identity, and residual risk
2. "Make humans meaningfully accountable" (§2.2)Human accountability, the agent value chain, and human oversight
3. "Implement technical controls and processes" (§2.3)Controls for agents, tests before deployment, monitoring during deployment, and change management
4. "Enable end-user responsibility" (§2.4)Information for "users who interact with agents", training for "users who integrate agents into their work processes", and tradecraft

After anomalies during "implementation or monitoring", the MGF recommends that organizations "re-evaluate the earlier dimensions", "re-assessing and further bounding the risks if needed" (§2).

AI agents ​

In §1.1, the MGF for Agentic AI tells organizations that "there is no consensus on what defines an AI agent". The MGF for Agentic AI is about "agents built on generative AI models" (§1.1). For agent components, action-space and autonomy, the types of risk, and systemic and multi-agent risks, see Agentic AI components and risks.

Changes in Version 1.5 ​

The MGF for Agentic AI "was launched" at the World Economic Forum in January 2026 (IMDA press release, 20 May 2026, ¶1). Version 1.5 includes "industry feedback from over 60 organisations" (press release, ¶3). The MGF for Agentic AI gives these changes in "What's new in this version":

  • Agent components — adds controls, and logging and monitoring, as components for "safe and reliable performance" (§1.1.1).
  • Protocols — "Updated for newer protocols, especially for agentic commerce" (§1.1.1).
  • Systemic and multi-agent risks — adds this type of risk (§1.2.3).
  • The four dimensions — adds an IMDA case study about how to apply the MGF for Agentic AI to OpenClaw (§2).
  • "Assess and bound the risks upfront" — adds risk factors, for example system complexity and the use of solutions from third parties (§2.1.1).
  • "Make humans meaningfully accountable" — identifies platform providers as different from system providers or app developers in the agentic AI value chain (§2.2.1). It also adds more methods to decrease automation bias, for example "monitoring human override rates and response times" (§2.2.2).
  • "Implement technical controls and processes" — adds information about the types of controls and how to select them (§2.3.1). It also adds recommended procedures for change management (§2.3.3).
  • "Enable end-user responsibility" — adds more information about "loss of tradecraft" and its effect on business continuity (§2.4.3).
  • Case studies — adds case studies from organizations for each dimension.

The case studies show how some organizations apply the MGF for Agentic AI (press release, ¶2). A case study shows the decisions for one deployment. These decisions are examples, and the MGF for Agentic AI does not tell all organizations to make the same decisions.

Other Singapore documents also give guidance about AI agents, generative AI, or AI risk. These documents are different from the MGF for Agentic AI. The Cyber Security Agency of Singapore (CSA), the Personal Data Protection Commission (PDPC), IMDA, and the Monetary Authority of Singapore (MAS) give these documents. Each document identifies its effect:

DocumentDatesEffect of the document
CSA, Securing Agentic AI – An Addendum to the Guidelines and Companion Guide on Securing AI SystemsThe CSA page has the date 17 June 2026. The PDF has the date June 2026.The PDF is "intended for informational purposes only and is not mandatory, prescriptive nor exhaustive" (PDF page 2).
PDPC, Advisory Guidelines on Use of Personal Data in Generative AIThe Guidelines have the date 20 July 2026. The public consultation "has closed on 1 July 2026".The Guidelines "are advisory in nature, are not legally binding on the Commission or on any other party" (¶1.5). They "neither modify nor supplement in any way the legal effect and interpretation of any laws cited, including, but not limited to, the PDPA" (¶1.5). The PDPA is the Personal Data Protection Act 2012 (¶1.1).
IMDA, Transparency Guidelines for Generative AI ChatbotsThe PDF has no date.The Introduction tells deployers: "These guidelines are voluntary and do not impose obligations". Footnote 3 tells deployers that the guidelines "do not replace, modify or otherwise affect any obligations that may apply under other laws, regulations, or sector-specific requirements". They are for deployers of "generative AI chatbots" that are "external-facing" (Introduction, "Generative AI Chatbots as a Start", p. 6). They are also "a reference for sectoral regulators to extend transparency to other AI applications in different sectors" (Executive Summary, p. 3).
MAS, Safeguards for Agentic Finance at Runtime (SAFR)Version 1.0 has the date 3 July 2026. Version 1.1 has the date 9 October 2026.SAFR is "jointly developed by the financial industry under MAS' BuildFin.ai initiative" (MAS page). The white paper gives SAFR "as a reference approach" for a "runtime governance layer", and SAFR "does not constitute regulatory guidance or supervisory expectations" (Introduction).

For the MAS Guidelines on Artificial Intelligence Risk Management for Financial Institutions of 7 October 2026, see MAS FEAT.

To compare AI governance frameworks, see Comparison. For Controls that are in more than one Modulos Framework, see Control overlap.

Use the framework in Modulos ​

Modulos gives the MGF for Agentic AI as two Frameworks with the name Singapore MGF for Agentic AI: OFF-17 at Organization level and MFF-17 at AI System level. Modulos recommends OFF-17 for the organization AI System and MFF-17 for each AI System that has one or more agents. For the Requirement table and the recommended sequence of work, see Use the MGF for Agentic AI in Modulos.

Frequently asked questions about the MGF for Agentic AI ​

For the FAQ about the MGF for Agentic AI, see Frequently asked questions about the Singapore MGF for Agentic AI.

Sources ​

Disclaimer

This page gives general information. It is not legal advice.