Singapore's Model AI Governance Framework for Agentic AI: What It Requires

Why Agentic AI Requires a Different Governance Model

Earlier AI governance guidance, in Singapore and elsewhere, was built around generative and predictive systems that produce a single output in response to a single input. Oversight for those systems focused on prompt and output review: what did the model produce, and was it accurate or appropriate.

Agentic AI systems behave differently. An agent executes multi-step sequences of actions, invoking tools, calling APIs, and making intermediate decisions without a human reviewing each step. Governance guidance for agentic AI reflects that shift. The obligation is no longer limited to reviewing an output; it extends to demonstrating that every action an agent took, and every decision that led to it, occurred within approved boundaries.

The Three Core Obligations

Singapore's governance guidance for agentic AI organizes around three interconnected requirements. Each one has direct implications for how an organization designs, deploys, and monitors autonomous agents.

Requirement	What It Means in Practice
Accountability	A named internal owner for each agent's behavior and outcomes; documented ownership at the agent level, not just organization-wide.
Human Oversight	Defined checkpoints where agent actions require human approval before execution; evidence that those checkpoints are exercised, not bypassed.
Traceability	Complete records of agent actions, decisions, and tool calls; sufficient to reconstruct the sequence of steps an agent took and under whose authorization.
Risk Tiering	Controls scaled to the reversibility and potential impact of agent actions; high-impact or irreversible actions subject to stricter approval requirements.

Accountability: Who Answers for an Agent's Actions

A recurring requirement in agentic AI governance guidance is that accountability must be assignable to a specific internal role, not distributed informally across a development team. In practice, this means an organization must be able to answer, for any deployed agent, who owns its behavior, who approved its permissions, and who is responsible if it causes harm or a compliance breach.

This differs from generic AI accountability statements in that it requires documentation tied to individual agents or agent classes, not a single organization-wide AI policy. Where multiple agents operate across business units, accountability mapping needs to identify ownership at the agent level, including agents built on shared infrastructure or reused across teams.

Human Oversight: Designed Versus Demonstrated

Governance guidance for agentic AI treats human oversight as something that must be demonstrated in operation, not simply described in a design document. This distinction matters because agents can be designed with human-in-the-loop checkpoints that are never actually exercised in production, either because thresholds are set too high or because approval steps are bypassed under operational pressure.

Meeting this requirement in practice means defining, for each agent, which categories of action require human approval before execution, which can proceed autonomously, and what evidence exists that the approval step was followed rather than skipped. High-impact or irreversible actions, such as financial transactions, data deletion, or external communications, are the actions most likely to require an explicit approval step.

Traceability: Reconstructing What an Agent Did and Why

Traceability requirements for agentic AI extend beyond logging a model's final output. Because an agent's behavior consists of a sequence of tool calls and intermediate decisions, traceability means being able to reconstruct that sequence after the fact: which tool was called, with what parameters, at what time, under which agent identity, and with what outcome.

This is a materially different logging requirement than prompt-and-response logging used for generative AI, and it is the requirement most likely to expose gaps in existing AI monitoring. Many organizations log model outputs but not the actions an agent took to produce them. Meeting this requirement typically involves structured logging at the tool-call level, linked to a specific agent identity and session context.

Operationalizing These Requirements

The three obligations translate into specific technical and operational controls. Policy documentation alone does not satisfy them. The controls most directly implicated include:

• Agent-level identity and authentication, so that each agent's actions can be attributed to a distinct, auditable identity rather than a shared credential.
• Scoped, least-privilege permissions assigned per agent, reviewed when the agent's role or deployment context changes.
• Structured tool-call logging that captures the action taken, the parameters used, the time of execution, the agent identity, and the outcome.
• Defined escalation paths specifying which categories of action require human approval and the mechanism by which that approval is recorded.
• Accountability records mapping each deployed agent to a named internal owner who approved its permissions and is responsible for its behavior.

Evaluation Questions for Governance Leaders

Organizations assessing their readiness against this guidance should be able to answer the following questions with documented evidence, not verbal assurance:

• Can each AI agent be uniquely identified, authenticated, and permissioned independently of the human or system that invoked it?
• Do complete logs exist for every tool call or action an agent takes, not just its final output?
• Which categories of agent action currently execute without human review, and is that appropriate given their potential impact?
• Who is accountable internally if an autonomous agent causes harm or a compliance breach?
• Could current controls demonstrate to an external auditor that agent behavior stayed within defined operational boundaries?