Senior Minister of State for Digital Development and Information Tan Kiat How told Parliament on 7 October that the Government is running sandbox trials on agentic artificial intelligence (AI) before imposing binding safety rules on tools that can plan, call APIs, and execute workflows with minimal human clicks.
Agentic systems differ from chatbots that simply draft text. They can chain tasks—querying databases, filing forms, or triggering payments—within guardrails set by developers. Tan said MDDI needs hands-on experiments to understand failure modes in Singapore’s civil service environment, where data residency and sector regulations vary across health, transport, and finance agencies.
From guidelines to tests
Singapore’s Model AI Governance Framework already nudges organisations to document risk and monitor deployments. Tan argued that framework is necessary but insufficient when models gain autonomy. Trials underway explore logging every action an agent takes, requiring human approval above certain dollar thresholds, and isolating test environments from production citizen data.
He was answering questions from MP Charlene Chen (PAP-Tampines) about safety requirements as vendors market “digital coworkers” to ministries. Tan did not name specific agencies or vendors, but GovTech has previously piloted context layers that encode agency rules as machine-readable policies—an architectural hint that sandbox work is not purely theoretical.
Frontier models and access
While Tan focused on deployment inside government, Minister Josephine Teo addressed frontier labs in a written reply on 6 October. Teo said Singapore engages developers for information and model access where appropriate but will not make access a legal requirement, fearing rigid rules could reduce cooperation. She pointed to developer testing, independent research, and partnerships with overseas AI safety institutes as complementary inputs.
Workers’ Party MP Fadli Fawzi had asked whether Singapore would copy US or UK pre-deployment evaluation mandates. Teo’s answer aligns with Singapore’s voluntary-compliance tradition in tech regulation, even as the country endorsed an international call for stronger frontier safeguards.
High-risk uses on the horizon
Teo also said MDDI is studying whether essential services need extra measures—more rigorous testing, independently verifiable safety evidence, tighter deployment controls, and stronger oversight. That review could eventually touch banks’ customer-service bots, already supervised by the Monetary Authority of Singapore for accountability when third-party vendors fail.
For now, Tan’s message is procedural: experiment in sandboxes, measure what breaks, then codify requirements. Citizens may not see a new app on their phones next week, but the trials determine whether an agent can renew a passport status or reschedule a clinic appointment without a officer in the loop—and what happens when the agent hallucinates a wrong date.
Industry watch
Local startups selling agent orchestration to enterprises will parse Tan’s comments for procurement signals. If government demand requires audit trails and kill switches, private-sector buyers will likely inherit the same expectations. Until rules land, the action is in pilot rooms, not statutes.