TL;DR

Singapore’s MDA regulator has published its “Model AI Governance Framework for Agentic AI”, which makes an argument UK financial institutions will recognise: an ethics committee, a written policy and a human nominally in the loop do not constitute control of an autonomous system. What the framework asks for instead is technical constraint, pre-deployment evaluation and continuous monitoring — evidence rather than attestation.

Governance as an engineering problem

The document treats assurance as something that must be demonstrated on an ongoing basis, not signed off once. It asks organisations to bound risk before deployment against the autonomy level, the sensitivity of accessible data, system complexity and the deployment environment. It then calls for evaluation before release, monitoring in production, testing of how agents interact with external tools, scenario-based exercises, and mechanisms to detect and respond when behaviour departs from expectation.

The reason this reads as urgent is the deployment pace behind it. Lloyds Banking Group has said it will hire hundreds of specialists to build agentic capability. Deutsche Bank has talked about compressing delivery cycles. JPMorgan Chase’s position is that the more code machines generate, the more the engineering basics — testing and governance among them — matter.

The UK parallel is already live

This lands days after the Bank of England raised its own testing expectations for frontier AI in regulated firms, and weeks after Germany’s BaFin began monitoring AI inside banks. The direction of travel across three jurisdictions is the same, and it points away from documentary compliance. Firms operating in the EU carry DORA’s resilience-testing obligations regardless of where they are headquartered.

For quality engineering teams the practical shift is in what the job covers. Pre-deployment validation starts to resemble conventional testing at much higher complexity — not only whether the system works, but how an autonomous agent behaves across varying conditions. Continuous monitoring replaces periodic test cycles. Scenario and risk-based evaluation looks a great deal like the resilience and chaos engineering work already spreading through banking.

Looking forward

The framework’s underlying claim is that an agentic system cannot be certified once and left running, because its behaviour shifts with context, inputs and interaction with other systems. If supervisors accept that, the burden moves from producing a policy to producing evidence that an agent’s authority is bounded in practice — that it cannot exceed its permissions, reach data it should not, or act outside regulation. Testing stops being a release gate and becomes the mechanism by which a bank governs at all.