Generative AI produced content, and the legal questions concerned that content: who owns it, whether it infringes, whether the training data was lawfully obtained. Agentic AI does things. It books, orders, negotiates, executes. The legal questions change accordingly, and Singapore is among the first jurisdictions to address them in a published framework.

What was issued

The Infocomm Media Development Authority released version 1.5 of its Model AI Governance Framework for Agentic AI on 20 May 2026, with an update following on 5 June. Alongside it, IMDA published a discussion paper on Legal Responsibility for AI Agents in May 2026.

The institutional division is worth noting. The Ministry of Digital Development and Information sets national AI strategy and public sector digitalisation policy; IMDA, a statutory board under the ministry, develops the governance frameworks and technical tooling. The framework is principles-based and voluntary, and is explicitly described as a living document intended to evolve as deployment experience accumulates.

The attribution problem

The hard question the discussion paper engages is attribution. When an autonomous agent acts on a principal’s behalf and something goes wrong, who is responsible?

Existing law offers analogies but no clean fit. Agency law addresses acts by human agents with authority granted by a principal, but assumes the agent is a legal person capable of bearing duties. Product liability addresses defective goods, but an agent that performs exactly as designed and still produces a harmful outcome is not obviously defective. Contract law addresses formation, but an agent that concludes an agreement raises questions about authority and mistake that the doctrine was not written for.

Nor is it clear who among the possible defendants should bear the loss: the deploying organisation, the model developer, the integrator that connected the agent to a payment system, or the counterparty who dealt with it.

Why voluntary frameworks still matter

A voluntary, principles-based framework attracts obvious criticism. It creates no enforceable obligation and no remedy.

That criticism understates its effect. Singapore’s earlier Model AI Governance Framework, also voluntary, propagated through procurement requirements, sectoral regulator expectations and contractual terms until it functioned as an operative standard for firms trading with Singapore-regulated counterparties. Financial institutions in particular have been expected to be transparent with customers and regulators about AI use, including how it influences decisions.

Frameworks of this kind also shape litigation. When a court eventually assesses whether a deployer acted reasonably, a published national framework the deployer ignored becomes evidence.

What to do now

Find out where agents are already running. Procurement approvals, customer service resolution, transaction monitoring and scheduling are the common entry points, and agentic capability tends to arrive as a feature update to an existing tool rather than as a new system with its own governance review. Then establish what authority those agents hold: spending limits, the ability to bind the organisation contractually, access to production systems. Finally, read the insurance. Professional indemnity, cyber and errors-and-omissions policies were mostly underwritten on the assumption of human error, and few of them say anything about a machine acting on its own initiative.

Summary

IMDA published version 1.5 of its Model AI Governance Framework for Agentic AI in May 2026, updated in June, together with a discussion paper on legal responsibility for AI agents. The framework is voluntary and principles-based, but Singapore’s earlier AI governance work became an operative standard through procurement and regulatory expectation. Organisations deploying agents should map where they are running, what authority they hold, and whether existing insurance covers what they do.