01 — ARIA Concepts

What is ARIA?

ARIA (Asset Registry for Intelligent Agents) is a reference architecture for managing the full lifecycle of enterprise AI primitives — agents, skills, instructions, knowledge bases, and orchestration configurations.

Think of it as TOGAF for AI assets: a metamodel that defines what these things are, how they relate to each other, and how they’re governed.

The Three Layers

ARIA operates across three layers:

Metamodel — What are these things and how do they relate?
- Five entity types classified by OASF
- Relationship types with governance implications
- Lifecycle states with gated transitions
Marketplace — How do you publish, discover, and compose them?
- GitHub repos with OASF manifest sidecars
- OCI registry for content-addressed artifact storage
- CI/CD pipelines for validation and publishing
Governance — How do you classify, control, and audit them?
- Microsoft Purview for sensitivity labels and DLP
- OASF governance overlay for policy-as-code
- Sensitivity label inheritance through dependency graphs
Distribution — How do non-technical users consume them?
- ARIA Catalog API translates OCI artifacts into platform-native installs
- Claude Desktop Extensions panel with enterprise allowlist
- Cowork contextual discovery based on OASF domain matching
- Web portal for browsing, filtering, and requesting access
- Governance enforcement is invisible — blocked assets are never shown

The Five Entity Types

Entity	What it is	OASF classification
Agent	An autonomous AI unit with identity and lifecycle	OASF Record (primary)
Skill	A reusable capability an agent can invoke	OASF Skill annotation
Instruction	Behavioral configuration (prompts, guardrails)	OASF Module (prompt_bundle)
Knowledge	Data corpus for grounding and retrieval	OASF Module (knowledge_base)
Orchestration	Routing and composition logic	OASF Module (orchestration_config)

Key Relationships

Agent ──invokes──────→ Skill        (1:N)
Agent ──governed_by──→ Instruction  (1:N)
Agent ──grounded_in──→ Knowledge    (1:N)
Orchestration ──composed_by──→ Agent (1:N)

Each relationship carries a governance implication. For example, when an agent invokes a skill, it inherits the sensitivity ceiling of that skill.

The Sensitivity Inheritance Rule

An AI asset inherits the highest sensitivity classification of any asset it depends on.

If your agent accesses a “Public” skill and a “Confidential” knowledge base, the agent itself is classified as “Confidential.”

This is enforced: - At CI time — the oasf-validate GitHub Action checks that the agent’s declared sensitivity_tier ≥ max(dependency tiers) - At runtime — Purview DLP policies prevent under-classified agents from accessing sensitive knowledge - At install time — the aria audit command validates governance before pulling artifacts

Next Steps

→ 02 — Marketplace Setup