Summary: Singapore used the World Economic Forum’s Davos stage to make the announcement: on January 22, 2026, the Ministry for Digital Development and Information (MDDI) released the Model AI Governance Framework for Agentic AI, drafted by the Infocomm Media Development Authority (IMDA). First, teams building or buying agent products can recast the document as supply-chain audit questions: where data-access boundaries sit, which checkpoints need human sign-off, how a multi-agent system degrades when it fails.
Singapore used the World Economic Forum’s Davos stage to make the announcement: on January 22, 2026, the Ministry for Digital Development and Information (MDDI) released the Model AI Governance Framework for Agentic AI, drafted by the Infocomm Media Development Authority (IMDA). MDDI’s positioning is unambiguous — nothing before it had handed enterprises a complete governance guide written for agentic AI; this is the first. Not one more principles manifesto: it answers, concretely, what organizations now face — once agents plan tasks, operate tools, and deal with other agents on your behalf, how is governance actually done?
The date carries a message of its own. That very day, Korea’s AI Basic Act entered into force; one week earlier, Anthropic had laid out Claude’s constitution. Regulation and self-policing are both leaving the model layer for the behavior layer: attention now covers more than what a model says — it extends to what an agent actually does, and who carries the consequences.
Why Autonomous Agents Need Rules of Their Own
The definition the framework adopts is worded with restraint: agentic AI covers systems able to plan and act over a series of steps toward assigned objectives, running with some degree of independence. Being able to act is where the shape of risk begins to change.
Two fresh layers of risk get named. One grows out of planning and reasoning, tool use, and protocols: an agent that reads and writes files, calls APIs, and hooks into the outside world through protocols like MCP can spin a single slip into a full chain of misguided actions. The other belongs to multi-agent systems alone — errors escalating between agents, plus competition or coordination nobody intended. The usual LLM risk list (hallucination, bias, data leakage, adversarial prompt injection) still stands; it no longer covers enough ground alone.
Four Pillars Holding the Framework Up
The substance sits in four interlocking dimensions, reaching teams that develop agents themselves and those buying from outside vendors alike:
- Measure the risks, then set the fences: run a risk assessment over an agent’s data access and degree of autonomy, and cap autonomy, tool usage, and data access early in the design — not after an incident forces it
- Keep humans answerable: responsibility lands on identifiable people across the value chain and lifecycle; oversight must push back on automation bias, significant checkpoints keep human approval, and audits run on a schedule
- Let technical controls span the whole lifecycle: controls ready during development, pre-deployment testing that runs entire workflows up to the multi-agent system level, a staged release path, then continuous monitoring, incident reporting, and failsafe mechanisms
- End users carry responsibility too: tell them plainly what the agent can do, which data it can reach, and whom to contact when something breaks; pair it with education so foundational skills do not erode through dependency
The Line From AI Verify to Agent Governance
The framework has a pedigree. Singapore’s governance releases line up in an almost unbroken sequence: the National AI Strategy of 2019; the first Model AI Governance Framework, published that same year and refreshed in 2020; AI Verify, the testing toolkit IMDA and the PDPC built in 2022; a generative-AI edition in 2024; and now one written for agents alone. Each release aimed at the deployment pattern of its day.
IMDA runs it as a living document: public consultation stays open, and enterprises are invited to send real case studies. Usable guidance first, field lessons second — the opposite of the EU AI Act’s legislative route. Teams mapping products onto both markets will find the Singapore text turns into a checklist fastest.
What Builders and Product Teams Should Take Away
Three consequences land in practice. First, teams building or buying agent products can recast the document as supply-chain audit questions: where data-access boundaries sit, which checkpoints need human sign-off, how a multi-agent system degrades when it fails. Second, transparency now reaches past model internals into agent behavior — saying what an agent does and sees is joining publishing a model constitution as a standard market expectation. Third, Asia’s compliance map keeps densifying: a Korean statute in force plus Singapore’s soft-law framework makes “design governance before shipping” the entry price for the APAC enterprise market in 2026, not a bonus.