Singapore tightens agentic AI guidance for physical systems

Singapore’s Infocomm Media Development Authority published version 1.5 of its Model AI Governance Framework for Agentic AI on May 20. The guidance sets rules on access controls, monitoring and human approval for AI agents that plan, decide and act across multiple steps as they move into warehouses, delivery networks and public spaces.

The framework defines agentic AI as systems that can plan and execute sequences of actions to complete user-defined goals. Agents may update databases, write files, control devices or perform transactions. IMDA advises organizations to assess use cases by data access, external system access, level of autonomy and task complexity.

IMDA recommends applying least-privilege permissions, establishing standard operating procedures for agent workflows and creating mechanisms to take agents offline when they malfunction. The document calls for gradual rollouts, ongoing testing and telemetry-based monitoring rather than one-time certification.

Speakers at a recent summit in Singapore raised safety concerns when software errors have physical consequences. Ya-Qin Zhang, founding dean of the Institute for AI Industry Research at Tsinghua University, warned that failures in digital systems can be amplified in the physical world and affect transport systems, drones, logistics networks and critical infrastructure. He cautioned: “Any risk in the digital domain will be amplified in the physical domain, and the physical domain will have a physical consequence.”

The framework treats governance as an iterative process. Human oversight must be adapted because continuous review of all workflows is impractical at scale. IMDA recommends human approval at checkpoints for high-stakes or irreversible actions and for outlier behavior. The agency flags automation bias and alert fatigue as supervision risks and recommends auditing oversight with indicators such as human override rates and response times, backed by automated real-time monitoring.

Grab, which is piloting autonomous vehicles and delivery robots in the Punggol district, relies on extensive simulation, closed-course and open-course testing, staged rollouts that start with a few units, and monitoring systems to detect unexpected failures, according to the company’s chief technology officer, Suthen Thomas Paradatheth.

IMDA notes that responsibility for embodied AI can span model developers, robotics manufacturers, semiconductor suppliers and infrastructure operators. Organizations and humans remain accountable for agent actions, and the framework calls for clear responsibility across the agentic AI value chain. Assigning responsibility becomes harder when systems continue to adapt after deployment through software updates and operational telemetry.

Applied Materials’ chief technology officer, Om Nalamasu, pointed to the need for better sensors, energy efficiency, advanced packaging and computing architectures for large-scale robotics deployment and added that many robotics systems will require purpose-built designs for specific industrial environments. Zhao Yuli, chief strategy officer at Galbot, described China’s focus on testbeds and partnerships and noted deployments of humanoid robots in retail, warehouse and pharmaceutical operations, including autonomous stores.

Professor Yutaka Matsuo of the University of Tokyo referenced an “AI Association” project to collect large-scale robotics data and mentioned national efforts including an AI Safety Institute and the Hiroshima AI Process as part of collaborative work on governance standards with Singapore and other countries.

Financial and retail firms are testing agentic workflows under human oversight. JPMorgan has implemented AI tools in investment banking to support research and client engagement and plans to hire more AI specialists. Some banks are testing security-focused models to detect vulnerabilities. IMDA includes a case study from OCBC Bank showing a source-of-wealth tool that parses documents and drafts memos but requires human review at critical decision points and final validation by designated reviewers. Walmart has outlined plans for multiple “super agents” for shoppers, employees, suppliers and developers and already offers a generative shopping assistant.

IMDA frames the guidance around deployment: use simulation, telemetry and iterative testing to inform rollouts; limit agent privileges; set clear human checkpoints; and maintain continuous monitoring and post-deployment assurance to detect and correct unexpected behavior in physical environments.