IMDA AI Governance — Layered Reference for Agentic Systems

Skill version: 0.3.1 · Primary framework: MGF for Agentic AI v1.5 · Supplement: MGF for Generative AI (2024)

⚠ Reference material, not legal or regulatory advice

This skill is engineering reference material. It is not legal advice, not regulatory advice, and not an assessment or certification of any kind. The maintainers are not affiliated with, endorsed by, or speaking for the Infocomm Media Development Authority (IMDA) or any other Singapore government agency.

The Model AI Governance Framework (MGF) for Agentic AI is voluntary guidance. It describes emerging good practice; most of it is phrased as what organisations "should consider". This skill does not turn it into law, and alignment with it does not discharge obligations under statutes that do bind you (the PDPA, sector rules such as MAS notices, contract terms). Always (a) verify any section reference or quotation against the official IMDA publication, and (b) involve your organisation's risk, security, legal and compliance owners before treating a design as approved.

Source-text posture: this skill does not reproduce or republish the framework. It provides engineer-facing interpretation, short attributed quotations of operative phrases, and section references, with a pointer to the official source. See DISCLAIMER.md, including § Copyright in source materials.

By using this skill you accept the full disclaimer in DISCLAIMER.md, including the "use at your own risk" terms and the maintainers' zero liability for any decision taken in reliance on this content.

This skill helps engineers make sure that what they design and build — architecture, agent design, permissions, approval flows, tests, monitoring, rollout plans, user-facing disclosures, internal guidelines — lines up with IMDA's MGF for Agentic AI. The agentic framework is the primary bar. IMDA's earlier MGF for Generative AI is applied as a supplement for what sits under the agent: the model, the data it was tuned or grounded on, and the content it generates. It is organised by where in the system the expectation lands, not by framework section number, so an engineer can work from the thing in front of them (a tool definition, an approval dialog, a deploy plan) rather than from the PDF.

Two conventions run through the layer files, checklists and templates (framework notes in frameworks/ restate the framework, and say where they add interpretation):

• [MGF §x.y] marks something the agentic framework itself says, with the section it comes from. That is the alignment bar.
• [MGF-GenAI <dimension>, p.N] marks something the generative-AI framework says. That framework has no section numbers, so it is cited by dimension and page.
• [Practice] marks an engineering pattern this skill recommends to meet that bar. It is not an IMDA requirement and can be substituted by anything that achieves the same outcome.

Keep that distinction in every review you write. A reviewer who presents a [Practice] item as "IMDA requires X" is overstating the framework.

Step 1 — Establish scope and the project's governance config

On first use in a project, check for .ai-governance.json at the project root. If present, load the listed frameworks' frameworks/<code>/README.md, and read riskTier and agentRegistry before doing anything else. If absent, establish three things with the user — or, when nobody is available to ask (an automated review, a CI run, a reviewer working from documents), infer them from the material, mark them provisional, list what you'd have asked as open questions, and carry on:

1. Is this agentic? The framework targets systems that plan and take actions over multiple steps towards a goal, built on SLM / LLM / MLLM models [MGF §1.1]. A generative feature with no tools or actions (chat, summariser, RAG Q&A, content generation) is out of the agentic framework's scope: use the generative-AI supplement as the bar instead, starting from checklists/new-genai-feature.md.
2. What is the organisation's role in the value chain? Model developer, tooling provider, platform provider, system provider / app developer, deployer, or several [MGF §2.2.1]. A team building an agent and running it in-house is both system provider and deployer.
3. What is the agent's current risk tier? If nobody has assessed it: when designing, run checklists/new-agent.md section 1 first; when reviewing, set a provisional tier as checklists/design-review.md step 2 describes. Every later decision — how many approval checkpoints, how much testing, how gradual the rollout — is calibrated to it.

Then suggest creating .ai-governance.json so future sessions and the changed-file guardrail don't re-ask:

{
  "aiGovernance": {
    "frameworks": ["sg-mgf-agentic", "sg-mgf-genai"],
    "role": ["system-provider", "deployer"],
    "riskTier": "unassessed",
    "agentRegistry": "docs/agents/",
    "reviewPolicy": "warn"
  }
}

frameworks lists the primary framework first. Use ["sg-mgf-agentic", "sg-mgf-genai"] for agents and ["sg-mgf-genai"] for generative features that take no actions. riskTier is one of unassessed, low, medium, high (see layer 02). agentRegistry is the folder holding one AGENT_CARD.md per agent. If the project cannot take that file, record the same facts in the project's AGENTS.md, CLAUDE.md or equivalent instruction file.

Framework codes: sg-mgf-agentic (primary) and sg-mgf-genai (supplement). How they relate, and to the 2020 baseline, is in frameworks/_index.md.

Step 2 — Pick the right entry point

You're reviewing something that already exists — a design doc, architecture diagram, ADR, internal guideline, PR or running system — and want to know whether it is aligned:

You're starting work on something. Open the matching checklist:

You want depth on a specific layer. Open the matching layer file:

Layer files say how to build it. What the agentic framework expects, section by section, lives in the dimension files listed below; use framework-map.md to cite a section in a PR description or review. The generative-AI supplement is in frameworks/sg-mgf-genai/. Disclosure for the system as a whole goes in a SYSTEM_CARD.md alongside each agent's card.

Framework files

Open these directly rather than following links from one file to the next:

Templates: AGENT_CARD.md, SYSTEM_CARD.md, ALIGNMENT_REVIEW.md, and ai-governance-nudge.sh (changed-file reminder hook).

Load-bearing principles (commit to memory)

These recur across the framework and settle most design arguments. Each links to the section that carries it.

How to use the layer ↔ framework split

The pattern is design once against the layers, then check against the framework:

1. Establish scope, role and risk tier (Step 1).
2. Walk the relevant layers/ to plan or critique the implementation.
3. Walk frameworks/sg-mgf-agentic/dimensions/01–04 (and, for the model and data layer, the generative-AI supplement) and confirm each expectation that applies is met, partly met, not met, or not applicable — with evidence (a file, a config, a test, a dashboard), not intent.
4. Record the outcome where the next engineer will find it: the agent's AGENT_CARD.md, the PR description, or an alignment review.

When reviewing, report against what exists, not what the author meant. A design doc that says "the agent will only read" while the tool grants write scope is a finding. A guardrail that lives only in the system prompt, protecting an irreversible action, is a finding. "We'll add monitoring later" for an agent heading to production is a finding.

Interactions with other obligations

• Personal data. Agents that touch personal data also carry PDPA (or other data-protection) obligations, which are law, not guidance. The MGF points to the organisation's data privacy policies and explicit consent where needed [MGF §2.4.2]; it does not replace them. If the personal-data-protection skill is installed, run it alongside this one.
• Cybersecurity. The MGF defers to CSA's Draft Addendum on Securing Agentic AI and GovTech's Agentic Risk & Capability Framework for control catalogues [MGF §2.3.1]. This skill does not reproduce them.
• Copyright and licences. Training, fine-tuning and RAG content can carry copyright and licence terms. The generative-AI framework flags this but leaves it to law [MGF-GenAI Data, p.11]; involve legal before using content you don't have clear rights to.
• Sector rules. Financial services, healthcare and public-sector deployments typically have binding rules on outsourcing, model risk and technology risk that sit above this framework.

Framework version

frameworks/sg-mgf-agentic/README.md and frameworks/sg-mgf-genai/README.md record which version of each framework the content reflects, when it was last verified, and what to watch for. The framework calls itself a living document and is updated as practice develops; check IMDA for a newer version before relying on a section number. See the upstream CHANGELOG.md and pin to a tag if you need stability.