magnus919/agent-skills/ai-governance

AI Governance

AI governance is the system an organization uses to decide, before a model is built and while it runs, who is accountable for an AI system, what risk it is allowed to carry, what evidence must gate each lifecycle stage, and how the organization reports and audits that posture. This skill teaches an agent to reason about and operate that system: it is a methodology skill, not a tool manual and not legal or security advice.

Scope: What This Skill Owns

This is a prevention-and-operations methodology: it gives the agent frameworks, decision models, and controls to design and run governance, not a claim that a system is compliant or safe. For every engagement, record the operating model, the risk tier, the evidence that gated each stage, and the accountable owner of each accepted exception.

When To Use

Load this skill to answer "how should we govern this AI system?" — standing up or maturing a governance program, tiering use-case risk, designing the operating model and decision rights, reviewing an LLM/agent system for governance and safety gaps, mapping a regulation to a compliance/control plan, scoring governance maturity, or preparing board-level reporting.

Capability admission and authority decisions

For “may this capability enter or remain in our environment?”, read references/capability-admission-and-update.md and use templates/capability-admission-record.md. For “may this agent act?”, use the earned-autonomy path below. Admission approves a component at a recorded revision and configuration; it does not issue a runtime grant. Link the two records in one review when both decisions are needed. Reuse current evidence and existing authorized scope; request a new decision only for a material change or an unmet approval condition.

Earned autonomy

When setting agent promotion/demotion thresholds or reviewing earned autonomy, read references/earned-autonomy.md and use templates/earned-autonomy-decision.md. Support accountable humans making capability/environment/action-class decisions; the agent cannot grant itself authority. site-reliability-engineering supplies operational evidence and agent-production-operations implements the approved control plan. This five-level autonomy ladder is separate from the Six-Level Governance framework.

Agentic engineering decision rights

When deciding which software-factory decisions remain with humans and which an agent may execute, read references/agentic-engineering-governance.md and use templates/engineering-decision-rights.md. Reuse scoped grants; distinguish technical readiness, intent acceptance, residual-risk acceptance and production authority. Check whether reserved human review has enough evidence, capacity and ability to stop the action.

Protected control-plane decisions

When an agent could weaken its own supervision, load protected-control governance and use the protected-control registry. Monitoring, alert routes, audit sinks, authority ceilings, kill switches, evaluator oracles, and gate configuration are protected assets. The controlled agent cannot be sole approver or sole evidence source for changing them. Missing, degraded, redirected, or suppressed mandatory controls are control-integrity incidents, not ordinary task failures. Governance owns rights and bounded human exceptions; operations verifies live controls, security designs external enforcement, harnesses cover child workloads, and evals challenge the evidence.

Reference Files (load on demand, one per task)

Progressive disclosure: load only the reference relevant to the current question.

Templates (fillable)

Use these to turn the methodology into working artifacts.

Scripts

Executable, flag-driven, stdlib-only Python CLIs with tests. Both accept a JSON input path and emit deterministic output; --json prints one JSON object on stdout; --dry-run previews without changing anything. Exit 0 on success; the maturity scorer also exits 1 on a critical posture, and both scripts exit 1 on input errors.

Evaluation and Configuration

• Eval manifest: evals/evals.json holds the output-quality cases (operating

model design, use-case risk tiering, 6L-G and deployment-posture review, agentic security and privacy review, impact-assessment closure, LLM-app governance review, fairness/accountability review, regulatory compliance mapping, board governance reporting, and GxP/data-integrity governance) used to grade this skill.

• Configuration: pytest.ini overrides the repository's root coverage settings so

the subprocess-based skill tests run cleanly; do not add a second override.

• Entry points: this SKILL.md is the router; README.md is the

human-facing overview for people evaluating whether to install the skill.

When Not To Use