Singapore’s Infocomm Media Development Authority has updated its Model AI Governance Framework for Agentic AI, offering practical guidance on managing the risks of AI systems that can plan and carry out tasks with a degree of autonomy. The framework emphasises bounded access, testing and meaningful human accountability. It is guidance rather than binding legislation.
Singapore’s Infocomm Media Development Authority (“<span class="news-text_medium">IMDA</span>”) published an <a href="https://www.imda.gov.sg/resources/press-releases-factsheets-and-speeches/factsheets/2026/updated-model-ai-governance-framework-for-agentic-ai" target="_blank" class="news-text_link">updated Model AI Governance Framework for Agentic AI</a> on 20 May 2026, following the framework’s initial launch in January. The update forms part of Singapore’s wider work on responsible AI governance and addresses the particular challenges posed by agents that can take steps towards a goal rather than simply respond to individual prompts.
The framework recommends that organisations assess whether an agent is suitable for a proposed task before deployment. That assessment should consider the risks created by its autonomy, the tools it can use and any access it has to sensitive information. IMDA advises organisations to define the agent’s scope and place appropriate limits on its permissions, including the services and data it may access. It also points to safety and security testing before deployment, staged rollouts and continuing monitoring as ways to identify and address problems in operation.
IMDA recommends clear allocation of responsibility across the agent’s lifecycle and checkpoints at which a person must review or approve an action. Transparency and user training also matter: people using an agent should understand that they are interacting with AI, what it is authorised to do and how to assess its outputs.
The framework does not itself impose new statutory obligations. Its value lies in providing a practical structure for organisations designing or deploying agentic AI: decide what the system may do, restrict the resources it can reach, test its behaviour and ensure that people can oversee consequential actions. The guidance reflects a broader shift from assessing an AI system’s answers alone to examining the actions it can take and who remains accountable for them.
