IMDA
Confidence 0.70 · 1 source · last confirmed 2026-08-30
Singapore’s Infocomm Media Development Authority, operating under the Ministry of Digital Development and Information (MDDI). IMDA’s significance to this wiki is its Model AI Governance Framework series — voluntary reference frameworks published early enough to shape how other jurisdictions’ rules get written, a strategy Singapore has run since the original 2019 framework.
The Model AI Governance Framework for Agentic AI, announced at Davos on 22 January 2026 and updated 20 May 2026, is presented as the world’s first governance framework aimed specifically at systems that take actions rather than produce outputs. Its four pillars:
- Bound the risks upfront — select appropriate agentic use cases and place limits on agents’ powers
- Make humans meaningfully accountable — define significant checkpoints at which human approval is required
- Implement technical controls across the agent lifecycle — baseline testing, access controlled to whitelisted services
- Enable end-user responsibility — through transparency and education/training
The structurally interesting move is pillar 1’s ex ante posture: risk is bounded at use-case selection, before deployment, rather than detected through output testing afterwards. That is the right shape for systems whose failures are actions, and it is a different regime from the disclosure-and-testing frameworks that dominate non-agentic AI governance.
What makes it worth reading alongside the engineering literature is the independent convergence: nothing in the practitioner or academic material in this corpus cites the framework, yet the same control set — bounded powers, approval checkpoints, least privilege, auditability — appears in a solo founder’s production setup, a product executive’s review bot, and a human-factors study of delegation regret.