Singapore IMDA AI Verify v2.0
Agentic AI Is Now a Standalone Test Category

On July 8, 2026, Singapore's Infocomm Media Development Authority (IMDA) released the AI Verify Foundation Framework Version 2.0. Compared to the 2024 initial release, v2.0's most critical update is: Agentic AI is now a standalone test category, with dedicated multi-step task testing and capability evaluation protocols.

What Actually Changed in v2.0?

AI Verify was originally Singapore's AI governance testing framework, giving enterprises tools to assess whether AI systems are "responsible." The 2024 first edition covered traditional AI and GenAI systems. v2.0 introduces three major changes:

• Agentic AI as a Standalone Test Category — For the first time, AI agents capable of autonomous planning, reasoning, and action are evaluated as a separate dimension, distinct from traditional AI and GenAI. Agentic AI systems can no longer "pass through" traditional tests — they now have dedicated evaluation protocols.
• Multi-Step Task Testing — Agents typically execute multiple sequential steps to complete a goal. v2.0 introduces testing methodologies specifically designed for multi-step workflows, evaluating decision quality, error recovery capability, and safety at each step.
• Capability Evaluation Protocols — The framework defines standardized evaluation methods covering tool-calling accuracy, context understanding, autonomous decision boundaries, and human oversight mechanism effectiveness.

This Is Not an Isolated Event

AI Verify v2.0 is Singapore's second major regulatory action since the January 2026 release of the Model AI Governance Framework for Agentic AI. Taken together, the signal is clear:

• January's MGF provided governance guidance — telling enterprises "what they should do";
• July's AI Verify v2.0 provides testing methods — telling enterprises "how to prove they did it."

Combined, these two frameworks create a complete guidance-to-verification loop. Paired with the PDPC data transparency rules that took effect in late 2025, Singapore has effectively become Asia-Pacific's most complete AI regulatory market — not the strictest (the EU still holds that title), but the most structurally coherent and enterprise-actionable.

Real-World Impact for Enterprises

If you deploy AI agents in Singapore or Asia-Pacific — whether customer service agents, code generation agents, or automated workflow agents — AI Verify v2.0 means three things:

First, your agents need to "prove" predictable behavior. AI Verify v2.0's multi-step task testing requires enterprises to document input, reasoning process, and output at every decision node. This is not a one-time assessment — the framework recommends periodic retesting, as agent behavior may change with underlying model updates.

Second, your agents need clear autonomy boundaries. The capability evaluation protocols require enterprises to define the boundaries within which agents can operate autonomously, along with human oversight mechanisms within those boundaries. When agents exceed boundaries, clear fallback mechanisms and audit trails are required.

Third, your agents need runtime audit trails. To satisfy AI Verify v2.0 evaluation requirements, enterprises must record complete agent runtime behavior — including tool calls, context changes, and decision paths. This is not covered by development-stage testing — it requires continuous production monitoring.

From "Voluntary" to "Mandatory"

AI Verify is currently still a voluntary framework. But history shows Singapore's regulatory path follows a consistent pattern: voluntary guidance → industry adoption → semi-mandatory periodic reporting → legislation. AI Verify Foundation participating enterprises already include Google, Microsoft, AWS, and other global tech giants, alongside DBS Bank, Singapore Airlines, and other local leaders. When industry benchmarks adopt, the "voluntary" window for smaller enterprises is narrowing.

More importantly, IMDA has explicitly stated it is evaluating whether AI Verify v2.0 testing methods should be incorporated into future regulatory requirements. This means today's investment in AI Verify will directly translate into tomorrow's compliance readiness.

Singapore vs. the World: Why This Matters

The EU AI Act is the world's most comprehensive AI legislation, but its implementation is slow — high-risk AI system requirements don't fully take effect until 2027. Singapore takes a different approach: industry frameworks as foundation, rapid iteration, gradual path to legalization.

For multinational enterprises, Singapore's path is actually more instructive — because it more closely resembles most countries' regulatory reality: not one-step legislation, but progressively tightening industry standards. India's July 2026 Digital India Act draft, with its risk classification framework, visibly references both Singapore and EU models.

If your concern is "how to manage AI agent compliance in real business operations," Singapore's practice is more worth watching than EU legislation — the former tells you "how," the latter tells you "that you must."

OOMeta's Position

AI Verify v2.0's Agentic AI testing requirements essentially demand that enterprises establish auditable, monitorable, and controllable runtime governance capabilities. This is not a problem a "policy document" can solve — it requires technical infrastructure.

The OOMeta AI Governance Platform provides runtime logging, policy engine, and audit trail capabilities that directly map to AI Verify v2.0's multi-step task testing and capability evaluation requirements. We are evaluating the inclusion of AI Verify v2.0 compliance frameworks into our platform's preset policy templates.