Singapore Agentic AI Governance

Overview

Singapore's Infocomm Media Development Authority (IMDA) officially released the Updated Model AI Governance Framework for Agentic AI on May 20, 2026, marking a significant advancement in global AI regulation. This pioneering framework is specifically designed to address the unique challenges and opportunities presented by agentic artificial intelligence systems, which are distinguished by their capacity for autonomous planning, reasoning, and action. Unlike earlier forms of AI, agentic AI can interact dynamically with external systems and take proactive steps to complete complex tasks, thereby introducing a new suite of risks that necessitate specialized governance approaches. The framework serves as a comprehensive guide for organizations in Singapore that are developing or deploying agentic AI solutions, whether in-house or through third-party providers.

The Updated Model AI Governance Framework for Agentic AI builds upon Singapore's established leadership in AI governance, extending the principles and best practices outlined in previous model frameworks for traditional and generative AI. It aims to provide a structured overview of the potential risks associated with agentic AI, such as unintended or erroneous actions, unauthorized data access, and cascading system failures, while also highlighting emerging best practices for managing these risks effectively. The framework is structured around four core dimensions: assessing and bounding risks upfront, ensuring meaningful human accountability, implementing robust technical controls and processes throughout the agent lifecycle, and enabling end-user responsibility through transparency and training. By offering actionable guidance across these dimensions, IMDA seeks to foster a trusted environment that encourages the safe and responsible adoption of agentic AI, ultimately promoting innovation while safeguarding against potential harms.

Definitions

Central to the Updated Model AI Governance Framework is the concept of 'Agentic AI.' This term refers to advanced artificial intelligence systems that possess capabilities beyond mere data processing or content generation. Agentic AI systems are characterized by their ability to autonomously plan a sequence of actions, reason about their environment, execute those actions, and adapt to real-time feedback to achieve specific goals on behalf of humans. They can break down complex tasks into smaller subtasks, select appropriate tools or functions, and interact with other AI agents or external systems to accomplish their objectives. This expanded autonomy and interactive capacity differentiate agentic AI from earlier generative models that primarily produce text, images, or predictions, introducing novel risks related to their potential to initiate changes in digital or physical environments.

The framework itself is part of Singapore's broader 'Model AI Governance Framework' (MGF) initiative. The MGF is a series of non-binding guidance documents developed by the Infocomm Media Development Authority (IMDA) and the Personal Data Protection Commission (PDPC). Its genesis can be traced to efforts to articulate a common AI governance approach and a set of consistent definitions and principles for the responsible use of AI. The MGF aims to provide greater certainty to industry players, promote the adoption of AI, and ensure that regulatory imperatives are met by translating ethical principles into pragmatic measures that businesses can adopt. This includes guidance on internal governance structures, determining AI decision-making models, and operational management. The Updated Model AI Governance Framework for Agentic AI specifically extends these foundational principles to address the unique characteristics and risks of agentic systems.

Governance and Institutional Framework

The development and continuous updating of Singapore's Model AI Governance Framework, including the latest iteration for Agentic AI, is spearheaded by the Infocomm Media Development Authority (IMDA). IMDA plays a pivotal role in shaping Singapore's digital future, and its efforts in AI governance are central to ensuring that technological advancements are harnessed responsibly. The framework itself, while a non-binding guide, strongly advocates for robust internal governance structures within organizations deploying agentic AI. It emphasizes the critical need for clear roles and responsibilities to be allocated across the entire AI lifecycle, encompassing developers, deployers, operators, and end-users. This clear delineation of accountability is crucial for managing the complex interactions and potential impacts of autonomous AI agents.

A cornerstone of the framework's governance philosophy is the principle of meaningful human accountability and oversight. Despite the enhanced autonomy of agentic AI, the framework unequivocally states that humans must remain ultimately accountable for the agents' actions and outcomes. This necessitates the establishment of effective human oversight mechanisms, including significant checkpoints where human approval is required, particularly for actions that could have material real-world impacts. Organizations are encouraged to adapt their existing internal structures and processes to integrate values, risks, and responsibilities specific to algorithmic decision-making and agentic systems. Furthermore, the framework's broader context includes initiatives like the establishment of a National AI Council, announced in February 2026, which provides strategic direction for Singapore's overall AI agenda, reinforcing the government's commitment to comprehensive AI governance.

Key Focus Areas

The Updated Model AI Governance Framework for Agentic AI is structured around four key dimensions, providing a comprehensive approach to managing the risks and ensuring the responsible deployment of AI agents. The first dimension, 'Assess and bound risks upfront,' guides organizations in conducting use-case-specific risk assessments. This involves considering agentic-specific factors such as the agent's level of autonomy, its access to sensitive data, the breadth of its available data, and the reversibility of its actions. To mitigate identified risks, the framework recommends designing agents with inherent limitations, such as restricting tool access, controlling permissions, defining operational environments, and utilizing sandboxed environments. These measures act as primary safeguards against unintended or harmful actions by the AI agent.

The second dimension, 'Make humans meaningfully accountable,' underscores the imperative of human responsibility even as AI systems gain autonomy. It calls for clear roles and responsibilities to be defined across the entire agent lifecycle, ensuring that developers, deployers, operators, and end-users understand their part in the accountability chain. This dimension also stresses the importance of effective human-in-the-loop mechanisms, particularly for high-stakes or irreversible actions, and cautions against automation bias in supervisory roles. The third dimension, 'Implement technical controls and processes,' focuses on ensuring the safe and reliable operationalization of AI agents through technical measures applied throughout the agent lifecycle. This includes incorporating technical controls for new agentic components like planning and tool use during development, and conducting rigorous testing for baseline safety and reliability, including execution accuracy, policy adherence, and tool usage, before deployment. Finally, the fourth dimension, 'Enable end-user responsibility,' highlights the importance of transparency and user empowerment. Organizations are advised to make their AI policies known to users, allow for feedback mechanisms where possible, and ensure that communications about AI systems are clear and easy to understand, thereby fostering informed and responsible interaction with agentic AI.

Implementation Framework

The implementation framework for the Updated Model AI Governance Framework for Agentic AI is designed to be practical and adaptable, enabling organizations to integrate its principles into their existing operational structures and development lifecycles. Organizations are expected to adapt their internal structures and processes to account for the new risks and capabilities introduced by agentic AI. This involves a fundamental shift in how AI systems are conceived, developed, and deployed, moving beyond traditional AI governance considerations to address the unique aspects of autonomous agents. The framework provides guidance on both technical and non-technical measures, encouraging a holistic approach to responsible AI deployment. It emphasizes that the successful implementation hinges on a deep understanding of agentic AI risks and the adoption of emerging best practices to manage these risks effectively.

A key supporting tool for implementation is AI Verify, an AI governance testing framework and software toolkit developed by IMDA and the Personal Data Protection Commission (PDPC). AI Verify assists organizations in validating the performance of their AI systems against internationally recognized AI governance principles, which have been updated to include considerations for generative AI applications and are relevant for agentic AI. This toolkit helps organizations assess their alignment with principles such as transparency, explainability, reproducibility, safety, security, robustness, fairness, data governance, accountability, and human agency. By leveraging AI Verify, organizations can systematically evaluate their agentic AI solutions, ensuring they meet baseline safety and reliability standards. The framework also encourages organizations to contribute case studies on their agentic governance experiences, fostering a collaborative environment for learning and continuous improvement in implementation practices.

Monitoring and Evaluation

The Updated Model AI Governance Framework for Agentic AI is conceived as a dynamic, "living document," reflecting Singapore's adaptive approach to AI regulation in a rapidly evolving technological landscape. This inherent flexibility means that the framework is subject to ongoing monitoring and evaluation, with a clear intention for it to evolve in response to stakeholder feedback and real-world deployment experiences. IMDA actively invites organizations to contribute case studies on their experiences with agentic AI governance. This participatory approach is crucial for gathering practical insights into the efficacy of the framework's recommendations and identifying areas that may require refinement or expansion as agentic systems become more sophisticated and widely adopted.

While the framework does not prescribe a formal, centralized monitoring body for compliance in the same way a binding regulation might, its emphasis on internal governance structures and the use of tools like AI Verify implies a decentralized, self-assessment-driven approach to evaluation. Organizations are expected to establish clear roles and responsibilities for monitoring and managing risks associated with their agentic AI systems, integrating these practices into their standard operating procedures. The continuous feedback loop, facilitated by industry engagement and the collection of practical examples, ensures that the framework remains relevant and effective. This iterative process of monitoring and evaluation allows Singapore to maintain its leadership in AI governance by adapting its guidelines to address emerging challenges and opportunities in the agentic AI domain, fostering a culture of continuous improvement in responsible AI development and deployment.

Penalties, Liability, and Appeals

As a 'Model AI Governance Framework,' the document issued by IMDA serves primarily as a non-binding guide and a set of best practices rather than a prescriptive legal regulation that imposes direct penalties, liability, or formal appeal processes. Its purpose is to provide organizations with a structured approach to identifying and managing the risks associated with agentic AI, thereby promoting responsible development and deployment. Consequently, the framework itself does not stipulate specific fines, sanctions, or criminal penalties for non-compliance. Instead, it aims to foster a proactive and responsible approach to AI governance within organizations, encouraging self-regulation and adherence to ethical principles.

However, the framework places significant emphasis on 'meaningful human accountability' for the actions of AI agents. This principle implies that while AI agents may act autonomously, ultimate responsibility and potential liability for their outcomes remain with the human developers, deployers, and operators. Organizations are expected to define clear chains of accountability across the AI lifecycle and establish human oversight mechanisms that can effectively intervene, override, or review agentic AI actions, especially those with real-world material impact. While the framework does not create new legal liabilities, organizations failing to implement the recommended governance practices might face existing legal consequences under other applicable laws (e.g., data protection, consumer protection, or tort law) if their AI agents cause harm. The framework encourages organizations to put in place internal processes for managing incidents and addressing potential harms, aligning with broader principles of good corporate governance and risk management.

Relationship to Other Instruments

The Updated Model AI Governance Framework for Agentic AI is not a standalone document but is intricately woven into Singapore's broader suite of AI governance instruments. It explicitly builds upon and expands the principles established in Singapore's earlier Model AI Governance Frameworks. The first edition of the Model AI Governance Framework was released in 2019, followed by a second edition in 2020, which focused on traditional AI systems. More recently, in 2024, IMDA and the AI Verify Foundation published the Model AI Governance Framework for Generative AI, addressing the unique challenges posed by generative models. The Agentic AI framework extends these previous efforts by specifically tackling the novel risks and complexities introduced when AI systems gain autonomous planning and action capabilities.

This iterative development ensures a consistent and evolving approach to AI governance that adapts to technological advancements. The framework maintains a sector- and technology-agnostic approach, allowing it to complement existing sector-specific requirements and guidelines that may be in place for various industries. Furthermore, the framework is closely related to AI Verify, Singapore's AI governance testing framework and software toolkit. AI Verify helps organizations assess their AI systems against established governance principles, and its capabilities have been updated to include considerations for generative AI applications, making it a relevant tool for evaluating agentic AI as well. This integrated approach ensures that Singapore's AI governance ecosystem is comprehensive, providing both high-level policy guidance and practical tools for implementation and assessment.

International Alignment

Singapore's Updated Model AI Governance Framework for Agentic AI reflects a strong commitment to international alignment and collaboration in the evolving landscape of AI governance. The framework explicitly states its alignment with international AI standards and principles, particularly those put forth by leading global bodies such as the European Union (EU), the United States (US), and the Organisation for Economic Co-operation and Development (OECD). This deliberate harmonization ensures that Singapore's approach is not isolated but contributes to and benefits from global conversations and best practices in responsible AI development and deployment.

Singapore has consistently positioned itself as a leader in global AI governance efforts, actively participating in international dialogues and initiatives. The development of this framework for agentic AI is seen as a timely response to the next phase of AI deployment, filling a critical governance gap on how to safely and accountably manage AI that not only processes information but also takes action. By offering practical guardrails across risk bounding, human accountability, technical controls, and end-user responsibility, the framework contributes to a growing regional and international convergence on key AI governance principles. This proactive engagement and alignment with global standards aim to foster cross-border cooperation, facilitate mutual recognition of trusted AI practices, and promote a consistent understanding of responsible AI deployment worldwide.

Implementation Timeline

Compliance Checklist

Sources and References