انتقل إلى المحتوى
العودة إلى المدونة
4 دقيقة قراءةبالإنجليزية

Singapore's Governance Framework for Agentic AI

Singapore's IMDA has issued the first governance framework built for agentic AI. Here is what it asks of organisations that deploy autonomous agents.

  • agentic AI
  • AI governance
  • Singapore
  • enterprise AI
  • AI policy

Singapore has released the world's first governance framework written specifically for agentic AI. Minister Josephine Teo announced it at the World Economic Forum in Davos on 22 January 2026. It covers AI systems that can reason, plan and act on their own, and gives concrete guidance to organisations that run them in production.

Singapore IMDA discusses agentic AI governance at Davos 2026
Singapore IMDA discusses agentic AI governance at Davos 2026

Why agents need their own rules

Earlier AI governance frameworks assumed systems that analyse data and make recommendations. Agents can start tasks, change databases, run multi-step workflows and coordinate with other agents without a human approving each step.

The Infocomm Media Development Authority (IMDA), which developed the framework, names four risks specific to these systems:

  • Unauthorised operations: agents acting beyond their intended scope
  • Cascading failures: errors in one agent spreading through connected systems
  • Information disclosure: agents with data access leaking sensitive information
  • Automation bias: people trusting agent decisions too much and reducing real oversight

The four parts of the framework

Assess and bound risks upfront. Before deployment, evaluate what the framework calls the "autonomy surface area": how much autonomy the agent has relative to the sensitivity of the task, which external tools and systems it can reach, what data it can read or change, and what happens downstream if it makes a mistake. Keep autonomy and tool access to the minimum the task needs, with sandboxes and permission systems constraining agents by default.

Make humans meaningfully accountable. As agents become more autonomous, responsibility can become diffuse. The framework asks organisations to define clear roles across the agent lifecycle and set explicit approval checkpoints. Having a human nominally responsible is not enough. The framework warns against supervisors rubber-stamping agent decisions, and expects decision points where people actively evaluate what the agent did.

Implement technical controls. These include sandboxing, safety testing of "tool usage, policy compliance, and workflow reliability" before production, continuous monitoring for anomalies in agent behaviour, and protection against privilege escalation. The framework advises a gradual rollout, so real-world behaviour can be observed and corrected before scaling.

Enable end-user responsibility. Users should get training on agent permissions and capabilities, clear information about what the agent can and cannot do, and a way to intervene or disable it. Users who understand an agent's limits are less likely to over-rely on it or miss cases that need human judgement.

Two further points matter for anyone building agents. Many production systems now use several specialised agents working in parallel, which is efficient but compounds risk when errors cascade from one agent to the next, so assess how one agent's failure would affect the rest. And the testing the framework describes is broader than model evaluation: it covers the agent's whole operating context, including how it uses external tools and handles edge cases.

Where it fits for UAE organisations

The framework is non-binding for now, but IMDA explicitly designed it to inform future regulation. It also sits within a regional trend. South Korea's AI Basic Act (2026) and Taiwan's AI Basic Act (2025) also address human oversight and data safeguards for autonomous systems, and a shared regulatory language built on human accountability and technical controls is forming across Asia-Pacific. For organisations working in several jurisdictions, that makes compliance planning more predictable.

The UAE has not yet issued governance guidance specific to agentic AI. Singapore's principles align well with the UAE's existing approach to AI regulation, which emphasises responsible development while supporting innovation, so they are a sensible reference for UAE teams putting agents into production now.

If you already run agents, I recommend one step this quarter: audit each one. List its level of autonomy, the tools it can call and the data it can read or modify, and tighten anything wider than the task needs. Then write down who reviews its decisions and how. That record improves oversight today and gives you a head start if guidance like this becomes mandatory.