Artificial intelligence is no longer just a tool that responds to prompts; it is increasingly an autonomous actor that plans, decides, and executes tasks on its own. This evolution, known as agentic AI, is reshaping how businesses operate, and Singapore is at the forefront of governing it responsibly. On 20 May 2026, the Infocomm Media Development Authority announced that the Model AI Governance Framework for Agentic AI ("Agentic MGF") has been updated to include more than ten case studies of real-world agentic deployments that provide practical illustrations on how organisations can operationalise different dimensions of the framework's recommendations.
What Is Agentic AI and Why Does It Matter?
Agentic AI systems are software systems, built from one or more AI agents, that can plan across multiple steps to achieve specified objectives. Unlike generative AI, which requires a prompt to produce content, agentic AI can take actions, adapt to new information, and interact with other agents and systems to complete tasks autonomously.
In practical terms, this means AI agents can already handle coding, customer service, financial analysis, and HR workflows with minimal human intervention. While agents are already transforming workplaces through coding assistants, customer service agents, and enterprise productivity workflows, their greater capabilities bring forth new risks. Some risks are already known to us and managed, such as the risks arising from the language models the agentic system is built on. However, these existing risks can manifest themselves differently, or compound themselves, due to the added capabilities of AI agents.
The Model AI Governance Framework for Agentic AI (Version 1.5) identifies five key risk categories businesses must be aware of: erroneous actions, unauthorised actions, biased or unfair actions, data breaches, and disruption to connected systems. The stakes are real: a hallucination in one agent can cascade into flawed downstream decisions across an entire automated workflow.
The Four Dimensions of the Framework
The Agentic MGF organises its guidance around four practical dimensions that any organisation deploying agentic AI, whether building in-house or using third-party solutions, must address.
Assessing and Bounding Risks Upfront
Before deploying an AI agent, organisations must rigorously assess whether the use case is even appropriate for agentic AI. The framework evaluates risk as a function of both the severity of potential harm and the likelihood of it occurring. Factors such as the reversibility of agent actions, access to sensitive data, and the agent's level of autonomy all feed into this assessment.
A compelling real-world illustration comes from Dayos, a Singapore-headquartered enterprise AI automation company featured in the updated Agentic MGF. Dayos replaced its own ServiceNow instance with an AI-powered ticketing agent. Every type of IT ticket was scored against three questions: severity of impact if the agent gets it wrong, reversibility of the action, and feasibility of human oversight at each step. The result was a tiered system, fully automated for low-risk tasks like password resets, human-approved for moderate-risk diagnostics, and entirely off-limits to the agent for high-severity actions like permission changes. This tiered approach is a model other organisations can adapt across industries.
Making Humans Meaningfully Accountable
One of the framework's most important emphases is that deploying an autonomous agent does not dissolve human responsibility. The organisations that deploy agents and the humans who oversee them remain accountable for the agents' actions. It can be challenging to fulfil this accountability when agent actions emerge dynamically and adaptively from interactions instead of fixed logic.
The updated framework addresses this through the case study of Tencent's CodeBuddy, an agentic coding system used internally by Tencent Cloud engineers. CodeBuddy employs a mix of preset secure defaults and configurable permissions to allow for meaningful human oversight without overly fatiguing the user. Complex commands are explained in plain English, aiding the user in making informed decisions. As an additional safety net, continuous and automated real-time monitoring takes place, and suspicious commands that are identified will still require human approval, even if the command has been pre-approved previously.
They have also defined significant checkpoints for human intervention: For example, the system defines by default which actions need human approval, such as editing files, running shell commands, making network requests, or using external tools.
The framework also warns against automation bias, the tendency to over-trust AI simply because it has performed reliably in the past. Organisations are advised to track metrics like human override rates and response times during agent reviews, as unusually low or fast approval rates may signal that human oversight has become superficial rather than substantive.
Implementing Technical Controls and Processes
Good governance is not just a matter of policy; it must be embedded in the technical architecture of agentic systems. The Agentic MGF recommends a lifecycle approach: design-stage controls, pre-deployment testing, and continuous post-deployment monitoring.
GovTech, the Government Technology Agency of Singapore, adopted a phased approach for rolling out agentic coding assistants. In the first phase, the rollout was limited to GovTech employees only, restricted to low-risk systems, and external tools such as MCP servers were not permitted. This contained any potential risks while simultaneously allowing GovTech to build the infrastructure, including central logging, monitoring, and a framework for connecting to approved external tools, needed for a broader rollout. The learnings from that first phase, including resolving network configuration issues and reducing the cognitive burden on human approvers, directly improved the second phase of deployment. It is a practical playbook that any organisation scaling agentic AI can follow.
Enabling End-User Responsibility
Governance does not end with developers and IT teams. The employees and customers who interact with AI agents daily are also part of the accountability chain. Ultimately, end users are the ones who use and rely on agents, and human accountability also extends to these users. Organisations should provide sufficient information to end users to promote trust and enable responsible use.
Workday, the global enterprise HR and finance platform, demonstrates this well. The user interface of Workday's AI agents clearly provides notice to users that the tool is powered by AI and is intended to support a select range of actions. For the purposes of the case study, two agents were highlighted: the Recruiter Agent, which supports the screening and evaluation of candidates for job roles, and the Conversational Scheduling Agent, which manages the full interview, orientation, and event-session scheduling lifecycle.
When these agents provide recommendations in sensitive workflows, they accompany each recommendation with an explanation of their reasoning, including the data considered, key factors that influenced the recommendation, and any noted uncertainties or risks. This transparency is not just good practice, it is a prerequisite for responsible human decision-making in an AI-augmented workplace.
What Should Businesses Do Now?
While the Agentic MGF does not impose binding legal obligations, it provides a clear picture of Singapore's regulatory direction for agentic AI and establishes industry best practices. Organisations should begin by conducting a gap analysis against the framework's four dimensions, reviewing contracts with agentic AI vendors to address security, performance, and liability, establishing internal governance structures with clear role assignments, and developing user policies with adequate training.
The updated Agentic MGF offers governance guidance grounded in real deployment experience rather than theory alone. For businesses that want to harness the transformative power of agentic AI while managing its very real risks, engaging early with this framework is not just prudent; it is a competitive advantage.