Government developing common oversight system to track AI agents and manage risks

As artificial intelligence (AI) evolves from helping public officers with routine work to carrying out tasks on their behalf, the government is developing a common oversight system to help agencies track AI agents’ actions, detect anomalies promptly and reduce the associated risks.

Senior Minister of State for Digital Development and Information Tan Kiat How said the authorities were developing a shared technology platform called GovHarness. He was delivering the opening address on Thursday (5 November), the second day of the Government Technology Agency’s STACK technology conference.

He noted that AI had mainly played a supporting role over the past few years, such as drafting emails and writing code. Emerging agentic AI can now plan tasks independently, use different tools and connect to other systems to complete more complex work.

However, as AI gains the ability to act autonomously, the risks also increase. He said, for example, that the consequences of AI drafting an email containing errors were very different from those of it sending the erroneous email on its own. Similarly, recommending a payment and actually making it were two different things.

“We must therefore be clear about which actions an agent is authorised to take, which tools it may use and when a person must approve its actions.”

GovHarness will incorporate common safeguards and testing mechanisms, removing the need for government agencies to build their own monitoring and testing systems.

It will also keep records of AI agents’ actions so that, when a system fails, developers can trace what an agent actually did, what information it used, which steps it took and where things went wrong.

Besides developing GovHarness, the Government Technology Agency (GovTech) has drawn up an Agentic Risk and Capability Framework to establish appropriate safeguards based on AI agents’ capabilities and the risks of the tasks they perform.

For example, higher-risk actions involving financial transactions, sending official communications or deleting data will require stricter controls, including human approval where necessary.

As more autonomous task-execution features are added to Desk, GovTech’s digital assistant for public officers, these guidelines will also be used to determine which actions AI agents can carry out automatically and which require human approval.

Meanwhile, as AI enables public officers without traditional software engineering knowledge to develop applications, GovTech is using shared platforms to help more officers innovate safely.

One such platform, Rabbit Deploy, allows public officers to deploy and test applications developed with AI assistance. It comes with built-in features such as government single sign-on, activity logs and continuous security scanning, so officers do not have to handle complex technical configurations themselves.

Rabbit Deploy currently has more than 8,000 users from 90 government agencies, with more than 5,000 applications running on the platform.

Tan, who is also Senior Minister of State for Health, cited the example of a finance officer who used AI to develop an application and deployed it through Rabbit Deploy. It cut the time spent consolidating information from multiple sources each month from 12 hours to one, saving more than 130 working hours a year.

Singapore has actively strengthened its governance of agentic AI in recent years. In January this year, the Infocomm Media Development Authority launched the Model AI Governance Framework for Agentic AI to guide businesses and organisations in limiting AI agents’ authority to act, strengthening human oversight and managing risks. The framework was further updated in May with real-world case studies and new safety guidelines to help businesses use the technology more safely.