Singapore's agentic AI governance framework, explained.

01What the framework is

Singapore's Infocomm Media Development Authority launched the framework at the World Economic Forum in Davos on 22 January 2026. It builds on the country's earlier model governance frameworks, and is addressed to any organisation looking to deploy agentic AI, whether it develops agents itself or uses third-party agentic solutions. Baker McKenzie described it as the world's first governance framework designed specifically for AI agents.

Version 1.5, published on 20 May 2026, folds in feedback from more than 60 companies. It adds controls, logging and monitoring to the core components of an agent, and covers the risks of many agents working together. It names third-party solutions as a risk factor, and adds practices against automation bias, such as watching how often people override an agent. IMDA calls it a living document.

02The four dimensions, in plain words

1. Assess and bound the risks upfront. Decide what each agent is for, then limit what can go wrong through design choices: limits on its autonomy, on the tools it can use, and on the data it can reach.
2. Make humans meaningfully accountable. The organisation that deploys an agent, and the people who oversee it, remain accountable for what it does. Put human approvals at significant checkpoints, and check regularly that those approvals are real rather than reflexive.
3. Implement technical controls and processes. Give each agent an identity and access controls, log and monitor what it does, test it before it is deployed, and manage changes to it.
4. Enable end-user responsibility. The people who use an agent, or are affected by it, should know what it is, what it can do and how to raise a problem.

03The detail that matters most: identity

The framework names agent sprawl as a risk: agents created and deployed across an organisation without central management. Its answer is identity. Every agent should have one, tied to a supervising agent, a human user or a department. All agent identities and their permissions should be issued from and tracked by a central system, so the organisation can see what it has deployed, spot anomalies and remove identities it no longer needs.

That is an inventory with owners, described as a control rather than a spreadsheet.

04What it asks of a deployer, as a checklist

1. Know which agents you run, what each may do, and what it can reach.
2. Bound each agent's autonomy and access to what its task needs, before it runs.
3. Give each agent an identity tied to a person or team, issued and tracked centrally.
4. Name the points where a person must approve an action, and check how often those approvals change anything.
5. Log what each agent does, test it before deployment, and review it when its model or tools change.
6. Tell the people who use or are affected by an agent what it is and how to raise a problem.

05Why it matters outside Singapore

The framework is a model rather than law, in Singapore or elsewhere. Its practices are written for any deployer, and they line up with what financial regulators in the region are asking for. The Monetary Authority of Singapore consulted in November 2025 on Guidelines for AI Risk Management that cover AI agents and expect accurate, up-to-date AI inventories. In August 2026 it said the guidelines would be finalised soon. In Australia, APRA's April 2026 letter to industry asks for an inventory of AI use cases and controls over agentic workflows.