Singapore: Understanding Singapore’s new Model Framework for Agentic AI Governance

Evaluate each use case for agentic AI deployment 

Organisations should identify and assess risks, considering the likelihood of risks and severity of impact. Factors that should be considered include the:

• level of tolerance for error in the concerned domain and use case;
• agent’s access to sensitive data;
• agent’s access to external systems;
• reversibility of the agent’s actions;
• agent’s level of autonomy; and
• complexity of the task.

Use design limits to put guardrails around risks

Limits may be defined for the agent’s access to tools and systems, autonomy, and area of impact. Examples of some design limits that can be implemented by organisations include:

• identity management and access control measures to track individual agent behaviour and allocate accountability;
• linking an agent’s identity to a supervising agent, human user or organisational unit; and
• ensuring that the agent’s permissions do not exceed those of a human user.

Clearly allocate responsibilities within and outside the organisation 

Accountability for the decisions and outcomes of agents should be flowed down the value chain.

Within the organisation, responsibilities should be allocated across teams, e.g.:

• Board members and senior management should set high-level goals and permissions for agents, and manage the overall governance approach.
• Product teams should define the requirements of agents and ensure that agents are implemented responsibly throughout their lifecycle.
• Cybersecurity teams  should define baseline guardrails and regularly conduct red teaming.
• Users should comply with usage policies and attend required training.

Outside the organisation, organisations should:

• define obligations in contracts with third parties based on risk tolerance, leveraging clauses on security arrangements, performance guarantees, data protection and confidentiality;
• maintain security and control of AI agents, e.g. through scoped API keys, per-agent identity tokens and robust observability of tool calls and access history; and
• provide clear information to users on organisational accountability and user responsibilities.

Design for meaningful human oversight 

Organisations should implement a system for effective oversight, involving:

• defining significant checkpoints or action boundaries for human approvals, especially for high-stakes or irreversible actions;
• designing approval requests to be contextual and digestible;
• training humans to identify common issues with agent outputs;
• regularly auditing the effectiveness of human oversight; and
• complementing human oversight with real time monitoring and alert tools.

Apply technical controls during design and development 

Technical controls should be implemented to mitigate identified risks. Organisations are encouraged to refer to sample controls in the Cyber Security Agency of Singapore’s Draft Addendum on Securing Agentic AI and the Government Technology Agency of Singapore’s Agentic Risk and Capability Framework.

Test agents before deployment 

Organisations should continue to adopt best practices on software and LLM testing, and address agentic AI-specific concerns through testing that covers:

• new risks in overall task execution, policy compliance, tool calling and error response;
• agent workflows;
• multi-agent system collaboration;
• performance in real or realistic environments;
• repeated rounds of testing across varied datasets; and
• the use of multiple evaluation methods.

Continuously monitor and test agents during deployment 

Agents may be deployed on a gradual basis to reduce risk exposure. Post-deployment, reporting and failsafe mechanisms should be implemented to allow organisations to:

• stop agent workflow and escalate to a human supervisor upon failure detection;
• identify points of failure in an agent’s workflow and interaction with other agents; and
• regularly audit the system for effectiveness.

When setting up a monitoring system, organisations should consider monitoring objectives, alert thresholds and intervention protocols.

Focus on transparency for users who interact with agents

Users who interact with agents on behalf of the organisation (e.g. customer service or sales agents) should be informed of:

• user responsibilities, e.g. verifying information;
• the fact that they are interacting with agents;
• agents’ range of actions and decisions;
• how the agent collects, stores and uses data; and
• human contact points for escalation.

Focus on education for users who integrate agents into their work process 

Users who use agents as part of their workflow should be trained on:

• foundational knowledge on agents, including use cases, prompting best practices, agents’ range of actions; and
• effective oversight of agents, including common issues with output.