← Global AI Laws & Governance Research Library
Singapore Model AI Governance Framework for Agentic AI
Prepared by Shubhank Suman · Last reviewed
CURRENT / VOLUNTARY NON-BINDING GUIDANCE
- Legal character
- AI governance framework
- Region
- Asia-Pacific
- First released
- January 2026
What it covers
The Model AI Governance Framework for Agentic AI is a specialised extension of Singapore's broader model-AI-governance approach for AI systems that can act with a degree of autonomy. An agent may plan and execute multi-step tasks, interact with external applications, access databases, invoke software tools, act for users, operate across multiple agents and continue after the initial instruction. The governance question is not only whether output is accurate, but what authority the agent has, what it can access and do, and who remains accountable. The framework covers assessing and bounding risks upfront, meaningful human accountability, technical controls throughout the agent lifecycle, and end-user responsibility through transparency and training. Version 1.5 adds operational examples and practices for multi-agent systems, third-party agents and automation bias.
Why it matters
Agentic systems can use credentials, tools, data and external services to take consequential actions. The framework translates autonomy into governance questions about permission boundaries, approval, traceability, monitoring and accountability.
Refer to it when
Use it when developing, deploying or procuring autonomous or semi-autonomous agents; setting tool and data permissions; designing human approval; governing multi-agent or third-party-agent systems; and addressing automation bias.
AI, privacy and cybersecurity relevance
- AI: Very High
- Privacy: High
- Cyber: Very High
AI
VERY HIGH RELEVANCE
The framework directly governs agentic AI through risk bounding, least necessary autonomy, human accountability, lifecycle controls, multi-agent governance, third-party-agent controls and protection against automation bias.
Privacy
HIGH RELEVANCE
The framework is not a privacy statute and does not replace Singapore's Personal Data Protection Act. It is relevant where agents access customer, employee, financial, communications, calendar, document or internal-system data. Organisations should limit access, assess necessity, prevent disclosure, govern third-party processing and retain human control.
Cybersecurity
VERY HIGH RELEVANCE
Agents may hold credentials, API, database, file-system, tool and communication access. Relevant controls include identity, authentication, authorisation, least privilege, segmentation, logging, monitoring, action approval, incident response, third-party risk, prompt-injection defence and tool-abuse controls.
Applicability
Voluntary guidance for organisations developing, deploying or procuring agentic AI, especially autonomous or semi-autonomous agents used internally or through third parties.
Current validity
CURRENT: VERSION 1.5, MAY 2026. Current voluntary AI governance framework; it is not an AI Act or statutory regulation.
Read first
Assessing and Bounding Risks → Human Accountability → Technical Controls and Processes → End-User Responsibility → Multi-Agent Systems → Third-Party Agents → Automation Bias.
Provisions / sections covered
Assessing and bounding risks
Determine appropriate agentic use cases, permitted decisions and actions, required human approvals, available tools and information, and prohibited actions. The governing principle is least necessary autonomy.
Meaningful human accountability
A human remains responsible where organisational accountability is appropriate, especially when an agent sends communications, makes transactions, modifies records, executes workflows or interacts externally.
Technical controls and lifecycle processes
Written policy is insufficient. Controls are needed across design, development, testing, deployment, operation, monitoring and post-deployment review.
Multi-agent systems
Distributed agent actions raise issues of identity, delegation, permission boundaries, inter-agent communications, traceability and responsibility for downstream actions.
Third-party agents
Organisations should understand access, tools, outbound data, monitoring and responsibility when using external agents.
Automation bias
Human oversight must be meaningful enough to identify and challenge problematic AI recommendations or actions.
Subject-matter tags
- Employment data
- Retention
- Surveillance
- AI governance
- Agentic AI
- Transparency
- Explainability
- Human oversight
- Risk management
- Accountability
Subject-matter tags describe what the indexed text discusses. They are not legal conclusions.
Related instruments in this library
Document page in the Knowledge Base - full text or PDF where it is available for this record.
This library locates primary and official sources. It does not interpret the law and it is not legal advice.