Live on AgentApproved

Singapore AI Governance Framework for Agents

The world's first governance framework built specifically for autonomous AI agents. Published 22 January 2026 by IMDA at the World Economic Forum in Davos. AgentApproved scores your agent against all 8 requirements across 4 governance dimensions.

What is the Singapore Model AI Governance Framework?

In January 2026, Singapore's Infocomm Media Development Authority (IMDA) published the Model AI Governance Framework for Agentic AI (v1.0) — the first governance framework anywhere in the world designed specifically for AI agents that plan, reason, and act autonomously.

This builds on Singapore's track record of AI governance leadership. IMDA published the original Model AI Governance Framework in 2019, updated it for generative AI in 2024, and now addresses the unique challenges of agentic systems — agents that initiate tasks, select tools, update databases, and adapt dynamically across multi-step workflows.

The framework is voluntary, not legislation. This is a deliberate choice. Singapore favours principles-based, pro-innovation governance backed by government-built testing tools (like the open-source AI Verify toolkit) rather than mandatory compliance regimes with penalties. The AI Verify Foundation, IMDA's wholly-owned subsidiary, co-develops these frameworks with 180+ member organisations including AWS, Google, IBM, Microsoft, and Salesforce.

The 4 Governance Dimensions

The Agentic AI framework identifies five core risk categories for autonomous agents: unauthorised actions, erroneous actions, biased or unfair actions, data breaches, and system disruption. It addresses these through four governance dimensions:

1. Assess and Bound Risks Upfront

Evaluate each use case by likelihood and impact. Restrict tool access to whitelisted services. Apply least-privilege permissions. Use sandboxed environments for testing. Define SOP-driven workflows agents must follow. Consider autonomy level, access to sensitive systems, and action reversibility.

2. Make People Meaningfully Accountable

Clear responsibility allocation across the entire lifecycle — developers, deployers, operators, end users. Human-in-the-loop mechanisms for high-stakes or irreversible actions. Override, intercept, and review capabilities. Protection against automation bias in supervisory roles.

3. Implement Technical Controls

Three lifecycle phases: design (tool guardrails, plan reflections, least-privilege), pre-deployment (workflow-level testing, policy compliance), and deployment (staged rollouts, real-time monitoring, anomaly detection, failsafe mechanisms). Continuous monitoring of high-risk actions.

4. Enable End-User Responsibility

Training and transparency on agent permissions. Differentiated approaches for internal vs external users. Clear escalation channels. User education to prevent over-reliance. Users as active stewards, not passive consumers.

Why This Matters for AI Agents

The Agentic AI framework addresses risks that simply don't exist in traditional or generative AI:

Cascading failures — one agent action triggers a chain of unintended consequences across systems
Tool misuse — agents selecting wrong tools or using them beyond intended scope
Accountability gaps — when an agent takes 15 autonomous steps, who is responsible for step 12?
Privilege escalation — agents gaining more access than intended through multi-step reasoning
Multi-step compounding errors — small mistakes compound across sequential autonomous decisions

Agent-specific requirements include agent identity and permissions systems (analogous to service accounts in IT), plan reflections where agents evaluate their own planned actions before execution, and workflow-level evaluations that test entire agent pipelines rather than just individual outputs.

How AgentApproved Maps to the Singapore MGF

AgentApproved's singapore-mgf scope maps your agent's runtime behaviour to all 8 requirements across the 4 dimensions. Each requirement is scored individually with an overall compliance grade.

Our SDK captures the evidence that proves compliance:

Dimension 1 (Risk Bounding) — agent identity tracking, tool usage logging, permission scope evidence
Dimension 2 (Accountability) — human oversight records, decision approval chains, reviewer attribution
Dimension 3 (Technical Controls) — automated event logging, session lifecycle tracking, anomaly patterns, pre/post deployment testing evidence
Dimension 4 (Transparency) — capability documentation, user-facing disclosure logging, escalation records

Every event is SHA-256 hash-chained and Ed25519 signed. The resulting attestation certificate is cryptographically verifiable — any party can independently confirm the compliance score without trusting AgentApproved.

Singapore MGF vs EU AI Act

Aspect	Singapore MGF	EU AI Act
Legal status	Voluntary framework	Binding legislation
Penalties	None AI-specific	Up to 7% global turnover
Approach	Principles-based, pro-innovation	Risk classification (prohibited/high/limited/minimal)
Agentic AI	Dedicated framework (Jan 2026)	Not yet specifically addressed
Testing tools	Government-built AI Verify (open source)	Conformity assessments by notified bodies
Interoperability	Mapped to NIST AI RMF + ISO 42001	Standalone

The frameworks are complementary, not competing. An agent attested against both the EU AI Act and Singapore MGF through AgentApproved demonstrates global compliance readiness. Use scope full to get scored against all frameworks simultaneously.

Why Singapore Matters

Singapore is the AI governance capital of the Asia-Pacific region:

S$1 billion government investment in AI R&D for 2025-2030
$1.31 billion in private AI funding attracted in H1 2025 alone
MAS financial AI guidelines — mandatory for all regulated financial institutions since December 2024
ASEAN AI governance leadership — published regional guidance adopted by member states
Bilateral agreements with the US, UK, and South Korea on AI governance standards
AI Verify Foundation — 180+ members including every major cloud provider

For any AI agent operating in APAC markets — especially financial services regulated by MAS — Singapore MGF compliance is becoming a baseline expectation. And because the framework is mapped to both NIST AI RMF and ISO/IEC 42001, compliance in Singapore reduces the effort needed for international standards.

Get Started

pip install agentapproved

Install the SDK, create an API key on the dashboard, and request your first Singapore MGF attestation with scope singapore-mgf.

PyPI Package →

Frequently Asked Questions

What is Singapore's Model AI Governance Framework for Agentic AI?

Published by Singapore's IMDA in January 2026, it's the world's first governance framework designed specifically for autonomous AI agents. It covers four dimensions — accountability, transparency, safety, and human oversight — with 8 specific requirements for agentic AI systems.

Is Singapore's AI governance framework mandatory?

Currently voluntary, unlike the EU AI Act which is binding legislation. However, Singapore's framework is increasingly referenced in procurement requirements and industry standards, making it a de facto requirement for organisations operating in APAC markets.

How does Singapore's framework compare to the EU AI Act?

Singapore takes a principles-based, pro-innovation approach; the EU AI Act uses risk classification with binding penalties up to 7% of global turnover. Singapore's framework is the only one specifically designed for agentic AI. The two are complementary — AgentApproved scores agents against both.

What are the four governance dimensions in Singapore's framework?

The four dimensions are: accountability (clear ownership and responsibility chains), transparency (capability disclosure and logging), safety (risk identification and mitigation), and human oversight (escalation paths and control mechanisms). AgentApproved maps evidence to all four dimensions automatically.

What is AI Verify and how does it relate to this framework?

AI Verify is Singapore's open-source AI testing toolkit, maintained by IMDA's AI Verify Foundation. It provides technical testing tools that complement the governance framework. AgentApproved provides runtime attestation that goes beyond point-in-time testing to continuous compliance evidence.

Does the framework apply to agents operating outside Singapore?

While not extraterritorial like the EU AI Act, the framework is rapidly becoming an APAC standard. ASEAN's Digital Economy Framework Agreement (DEFA) references Singapore's governance approach, meaning compliance with this framework positions you for broader ASEAN market access.