On January 22, 2026, Singapore’s Ministry for Digital Development and Information (MDDI) launched the Model AI Governance Framework for Agentic AI at the World Economic Forum in Davos, developed by the Infocomm Media Development Authority (IMDA). MDDI positions it as the first framework in the world to offer enterprises a comprehensive governance guide built specifically for agentic AI — not another statement of principles, but a direct answer to the question organizations now face: when agents plan, call tools, and coordinate with other agents on your behalf, what does governance actually look like?
The timing matters. Korea’s AI Basic Act came into force the same day, and Anthropic had published Claude’s constitution the week before. Regulation and self-discipline are both shifting from the model layer to the behavior layer: the market no longer cares only about what a model says, but about what an agent does — and who answers for it.
Why Agents Need Their Own Rulebook
The framework’s definition is deliberately modest: agentic AI systems plan and take action across multiple steps to achieve specified objectives, operating with some degree of independence. That capacity to act is exactly where the risk structure changes.
It names two layers of new risk. The first comes from planning and reasoning, tool use, and protocols: once an agent can read and write files, call APIs, and wire itself into the world through protocols, a single mistake can compound into an entire chain of wrong actions. The second is specific to multi-agent systems: escalation of errors between agents, and unintended competition or coordination among them. The familiar LLM risk list — hallucination, bias, data leakage, adversarial prompt injection — still applies, but it no longer suffices.
The Four Governance Pillars
The substance of the framework is organized around four interlocking dimensions, and it applies whether you build agents in-house or source them from third parties:
- Assess and bound risks upfront: run risk assessments over each agent’s data access and autonomy, and impose design limits on autonomy, tool usage, and data access early — not as an incident-response afterthought
- Make humans accountable: responsibility must be allocated to named owners across the value chain and lifecycle; oversight must be designed to counter automation bias, with human approvals at significant checkpoints and regular audits
- Apply technical controls across the lifecycle: controls during design and development, pre-deployment testing of entire workflows at the multi-agent system level, staged rollout, plus continuous monitoring, reporting, and failsafe mechanisms
- Assign end-user responsibility: tell users transparently what the agent can do, what data it touches, and who to escalate to; educate them while ensuring foundational skills don’t decay into dependency
From AI Verify to Agentic Governance
The framework did not appear from nowhere. Singapore’s governance lineage is remarkably consistent: the National AI Strategy in 2019, the original Model AI Governance Framework published that year and updated in 2020, the AI Verify testing toolkit from IMDA and the PDPC in 2022, the generative AI edition in 2024, and now an agent-specific framework. Each iteration targets the deployment pattern of its moment.
IMDA also frames it as a living document, open to public consultation, with an explicit invitation for enterprises to submit real case studies. This pattern — ship operational guidance first, harvest field feedback second — contrasts with the EU AI Act’s legislative route. For teams planning products across both markets, the Singapore document is the one you can convert into an audit checklist today.
What It Means for Builders
Three practical effects. First, if you are building or procuring agent products, this document converts directly into supply-chain audit questions: where are the data-access boundaries, which checkpoints require human sign-off, and how does a multi-agent failure degrade safely. Second, transparency is extending from model internals to agent behavior — disclosing what an agent does and sees is becoming the same class of market expectation as publishing a model constitution. Third, Asia’s compliance map is thickening: Korea’s statute in force plus Singapore’s soft-law framework means “design governance before shipping” is a 2026 entry requirement for the APAC enterprise market, not a bonus point.
Sources
- Singapore Launches New Model AI Governance Framework for Agentic AI — MDDI
- Model AI Governance Framework for Agentic AI — IMDA
- Launch of a new Model AI Governance Framework for Agentic AI — Stephenson Harwood
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.