IMDA and the AI Verify Foundation published the Starter Kit for Testing Agentic Systems for Safety and Reliability (Version 1.0) in October 2026, a practical reference for testing AI agents that 'act on their environment with limited human involvement' rather than only generating content. It lays out a four-step process — identify the relevant risks for testing, create realistic and robust tests, run tests and analyse logs, and continuously monitor and re-test — and builds on IMDA's earlier Starter Kit for Testing LLM-Based Applications and the Model AI Governance Framework for Agentic AI's third dimension (technical controls). It was released alongside the AI Verify Foundation's accreditation of its first five third-party AI testing firms (AI TAP) and a new Frontier AI Resilience and Security Taskforce (FAIRS).
OFFICIAL TEXT: imda.gov.sg ↗ · IN FORCE · IMDA SINGAPORE
| Document | IMDA/AI Verify Starter Kit for Testing Agentic Systems — Starter Kit for Testing Agentic Systems for Safety and Reliability |
| Issued by | Infocomm Media Development Authority (Singapore), with the AI Verify Foundation |
| Type | Framework |
| Status | In force |
| Published | Oct 14, 2026 |
| Applies to | Enterprise developers, testing teams, third-party testers and compliance/responsible-AI professionals conducting agentic-AI testing, including banks deploying AI agents; voluntary guidance from IMDA and the AI Verify Foundation |
| Official source | imda.gov.sg ↗ |
| Use cases | Generative & agentic AI · Model risk management · Third-party & vendor AI · AI governance (general) |
What are the key points of IMDA/AI Verify Starter Kit for Testing Agentic Systems?
- Four-step process: identify the relevant risks for testing (translating 'safety by design' and residual risks into test scenarios); create realistic and robust tests (test environment, logging, task/metric/evaluator design, including tool-use and handoff accuracy); run tests and analyse logs (repeated runs, model-harness swaps, context-window compaction risk); and continuously monitor and re-test, since agentic systems 'introduce additional sources of change through tool and service integrations, memory, and long-horizon interactions'.
- Includes five case studies: RFP extraction agents (Cohere), a code-fix agent (CyberSierra), SDLC-support agents (KBTG), a financial reconciliation agent (EarlyBird AI and Knovel Engineering), and a content-verification agent (IMDA's own MAGIC Agents).
- The financial-reconciliation case study — an agentic accounting system for small enterprises doing invoice extraction, bank-transaction classification, reconciliation and monthly close — tested stateless classification, stateful memory persistence across a three-tier client-memory/global-vendor-master/LLM-fallback architecture, and adversarial prompt-injection attacks via uploaded invoices, run through the Global AI Assurance Sandbox.
- Released the same day as the AI Verify Foundation's AI Testing Personnel Accreditation Programme (AI TAP), which accredited its first five firms — AIDX Tech, Knovel Engineering, PwC Singapore, Resaro and Vulcan — across five risk areas.
- IMDA also formed the Frontier AI Resilience and Security Taskforce (FAIRS, 20+ members), targeting a practical-guidelines report for the infocomm/media sector by H1 2027.
What did IMDA/AI Verify Starter Kit for Testing Agentic Systems change for banks?
The first IMDA/AI Verify testing methodology specifically for agentic systems, with a real financial-services case study (an invoice/reconciliation agent) — direct follow-through on the agentic-AI testing guidelines the tracker has been watching since the January 2026 agentic framework launch.
Has IMDA published testing guidance for AI agents?
Yes. In October 2026 IMDA and the AI Verify Foundation published the Starter Kit for Testing Agentic Systems for Safety and Reliability (Version 1.0), a four-step testing methodology for agentic AI.
Is there a case study on financial or accounting AI agents?
Yes. The starter kit includes a case study on EarlyBird AI's agentic accounting system, which tested invoice extraction, bank-transaction classification, reconciliation, stateful memory persistence and adversarial prompt-injection resistance.
| Date | Document | Status |
|---|---|---|
| Jan 22, 2026 | IMDA Model AI Governance Framework for Agentic AI — Model AI Governance Framework for Agentic AI | In force |
| May 30, 2024 | IMDA Model AI Governance Framework for Generative AI — Model AI Governance Framework for Generative AI: Fostering a Trusted Ecosystem | In force |
| Jun 7, 2023 | AI Verify Testing Framework — AI Verify Testing Framework and Toolkit (AI Verify Foundation) | In force |
| Nov 2, 2026 | HKMA discussion paper on banks' use of Agentic AI — Discussion Paper on the Use of Agentic A.I. by Banks | Final |
| Oct 30, 2026 | Machado interview with Bloomberg (Oct 2026) — Interview with Bloomberg — Pedro Machado | Final |
| Oct 28, 2026 | Selig Frontier Forum Remarks on Super Intelligence (Oct 2026) — Remarks at Frontier Forum on Super Intelligence | Final |
Follow every document these regulators publish
when one of these regulators moves, the next morning's brief says so · six sourced stories · 7 am PT · free
plus every regulation, bank and agent page update, the morning after · leave any morning