Glossary · term

Model AI Governance Framework for Agentic AI

The Model AI Governance Framework for Agentic AI is voluntary guidance from Singapore's Infocomm Media Development Authority for organizations that deploy AI agents, whether developed internally or supplied by a third party. Version 1.5 organizes its guidance into four dimensions: assess and bound risks upfront, make humans meaningfully accountable, implement technical controls and processes, and enable end-user responsibility. It is a named framework, not a generic synonym for agent governance.

Regulation2026-01-22Wave 3 · 2025–26Maturity: 3/5

Origin and context

IMDA launched version 1.0 on 22 January 2026, building on Singapore's 2020 Model AI Governance Framework. Version 1.5 followed on 20 May and was updated on 5 June after feedback from more than sixty companies. It retained the four-dimension structure while expanding treatment of multi-agent and third-party risks, control types, change management, automation bias, and contributed deployment case studies.

Sources: s1, s2, s3, s4

Why it matters

The framework shifts governance attention from model outputs alone to systems that plan and act through tools. It asks deployers to choose suitable use cases, limit permissions and action-space, allocate responsibility across the value chain, design meaningful approval points, test before and after deployment, monitor behavior, manage changes, and inform end users. Its value is a common review structure; the document does not prove that any listed control is sufficient.

Sources: s2, s3, s4, s5

Example

A company reviewing an agent that can read invoices and initiate payments can use version 1.5 to document the use case, impact and likelihood, data and tool access, transaction limits, escalation rules, responsible humans and suppliers, pre-deployment tests, logs, monitoring, change controls, and user disclosures. Calling that assessment `aligned with the MGF` should identify the version and evidence. It does not mean the system is certified, legally compliant, or safe in every context.

Sources: s2, s3

How it differs

Automation bias in agentic AI

Automation bias is one human-oversight risk addressed by version 1.5. The framework recommends practices around meaningful accountability, but the psychological and organizational phenomenon is broader than this instrument.

Agent delegation chain

Delegation chains describe how authority can move among agents. The MGF addresses multi-agent complexity, identities, permissions and responsibility, but it is a governance framework rather than a delegation protocol or formal authorization model.

Frontier Compliance Framework

A frontier-compliance framework is a provider-specific approach to high-consequence model development and incidents. IMDA's MGF is public voluntary guidance aimed primarily at organizations deploying agentic systems; neither is an alias for the other.

Maturity and evidence

Maturity is rated 3. The instrument has a formal publisher, two substantive versions, a stable organizing structure, documented external feedback, implementation examples, and independent professional analysis. It remains young and explicitly living. The reviewed evidence does not show standardized conformity assessment, binding adoption, broad longitudinal implementation, or measured effectiveness across sectors, so maturity 4 would be premature.

Sources: s1, s2, s3, s4, s5

Limits and open questions

The MGF is principles-based guidance, not law or certification. Organizations still need context-specific technical, legal, safety, privacy, employment, accessibility and sector review. Contributor case studies can illustrate practices without independently validating outcomes. Terms and recommendations can change with later versions, and controls such as approval prompts, logging, model-based safeguards or MCP filtering can fail or create new risks. Cite the exact version and do not transform recommendations into universal requirements or assurance claims.