Singapore is moving AI agents from experiment to public-sector infrastructure, but it is starting with a simple question: who is responsible when software acts on its own?
Singapore’s latest AI move is not another chatbot launch. It is a control system for the next stage of office automation, where AI agents can plan tasks, use tools and act at machine speed across government work.
As The Straits Times reported, GovTech is developing a registry of AI agents for about 150,000 public officers, giving the government a way to track who owns each agent and what it does. That sounds administrative, but it goes to the heart of the agentic AI problem. A chatbot can suggest a sentence. An agent can schedule a meeting, draft a report, touch a file, send a request or carry out a sequence of steps before a human has caught up.
The registry sits inside GovTech’s AI Assistant Desk, a suite being piloted with some public officers before a broader rollout later in 2026. GovTech chief executive Goh Wei Boon has framed the goal plainly: a secure personal digital assistant for every public officer. The important word is secure. Singapore is not treating AI adoption and AI governance as separate tracks. It is trying to build both at the same time.
Most countries are still talking about AI rules in broad terms. Singapore is turning the problem into something operational. If an agency deploys an AI agent, the government wants a record of the owner, the allowed activities and the safeguards around that agent. That makes AI use visible before it becomes sprawling and difficult to govern.
This matters because agentic systems do not behave like ordinary software. A traditional application follows a known workflow. An AI agent can decide the steps needed to complete a task, especially when it has access to tools, files, calendars, forms or internal systems. That flexibility is the value. It is also the risk.
GovTech’s guardrails show where the danger points are. The AI Assistant Desk is designed to apply custom rules and sanctioned AI tools, while keeping security layers in place even when third-party AI tools are added or replaced. Reported controls include blocking agents from deleting files, preventing external emails in certain settings, capping recipient numbers to avoid spam and running automated checks for offensive language before content enters or leaves AI systems.
These are not glamorous features. They are the difference between useful automation and blind delegation. In a private company, a poorly controlled agent might cause embarrassment or data leakage. In government, the stakes can include citizen data, public trust and the reliability of services people depend on.
Singapore Is Using Government As The Test Bed
The timing is not accidental. Singapore established a National AI Council in February 2026, chaired by Prime Minister Lawrence Wong, to give strategic direction to the country’s AI agenda. In May, the government released an update to its National AI Strategy, setting out refreshed priorities for AI adoption across public services, research, governance and industry.
The public service is already moving. More than half of Singapore’s 150,000 public officers regularly use Pair, the government’s AI chatbot, for productivity, writing and research. That gives Singapore a base of AI users before agents arrive at scale. It also means the government has learned the obvious lesson early: once employees find AI useful, usage spreads faster than policy manuals can keep up.
That is why a registry is more than an internal compliance tool. It is a way to avoid the messy pattern many organizations are now facing, where staff quietly stitch together AI tools, browser extensions and workflow agents without a clear view of what data is being touched or which system is taking action.
Singapore’s recent AI Agents Sandbox with Google, the Cyber Security Agency of Singapore, GovTech and IMDA points in the same direction. The sandbox studied computer-use agents in real public-service settings and found that today’s digital environment was built mainly for human users. Identity systems, permission controls and authentication frameworks may need to evolve before agents can operate safely at scale.
That is a practical insight. You cannot simply drop autonomous tools into systems designed around human judgment and hope policy fills the gap. The infrastructure has to know when an agent is acting, what it is allowed to do and when a human should step back in.
The Bigger Lesson For Business
Companies should pay attention because Singapore is dealing with a problem that will soon show up everywhere. AI agents are attractive because they promise to remove routine work from teams that are already stretched. Coding, reporting, scheduling, testing and risk checks are obvious early targets. But productivity gains will be limited if every deployment creates a new security exception.
GovTech is also developing an AI tool for automated penetration testing across about 2,000 government systems containing citizen data and transactions. That shows how Singapore is thinking about AI on both sides of the security equation. AI can create new risks, but it can also help inspect systems faster than manual processes allow.
The forward-looking question is whether Singapore’s registry model becomes a template for other governments and large companies. It may. As AI agents move from pilots into daily operations, leaders will need inventories, permissioning, logs and accountability before they can scale with confidence. The winners will not be the organizations that give every employee an agent first. They will be the ones that know exactly what those agents are doing.
Also read: AI designed vaccines are moving from theory into human trials • OpenAI makes ChatGPT memory more active and harder to ignore • Washington is testing a new ownership model for AI companies
