Categorical Agent Registry (CAR)

The Categorical Agent Registry is the universal certification and classification system for autonomous AI agents. CAR provides a standardized identifier format, domain taxonomy, capability levels, and certification tiers that enable agents to be registered, discovered, and trusted across organizational boundaries.

The Problem

As AI agents become autonomous actors in enterprise systems, organizations face critical gaps:

• Discovery: How do you find an agent with specific capabilities?
• Trust: How do you know an agent is safe to grant access?
• Routing: How do you direct tasks to the right agent?
• Governance: How do you enforce regulatory compliance?

The Solution: CAR String

CAR provides a compact, machine-parseable identifier that encodes an agent's identity, capabilities, and version:

a3i.vorion.banquet-advisor:FHC-L3@1.2.0
 │     │         │         │   │   └── Version (semver)
 │     │         │         │   └── Capability Level (Execute)
 │     │         │         └── Domains (Finance + Hospitality + Communications)
 │     │         └── Agent Class
 │     └── Organization
 └── Registry

Key design decision: Trust tier is not part of the CAR string — it is computed at runtime from certification status, behavioral history, and deployment context.

Effective Autonomy = MIN(CAR_Certification, Vorion_Runtime_Score)

Core Components

Capability Domains (10 codes)

Code	Domain	Description
A	Administration	System admin, user management
B	Business	Business logic, workflows, approvals
C	Communications	Email, messaging, notifications
D	Data	Data processing, analytics, reporting
E	External	Third-party integrations, APIs
F	Finance	Financial operations, payments, accounting
G	Governance	Policy, compliance, oversight
H	Hospitality	Venue, events, catering management
I	Infrastructure	Compute, storage, networking
S	Security	Authentication, authorization, audit

Capability Levels (8 levels)

Level	Name	Approval Model
L0	Observe	Every action (read-only)
L1	Advise	Every action (recommendations)
L2	Draft	Before commit (staging)
L3	Execute	Per action (human approves)
L4	Autonomous	Exception-based (policy bounds)
L5	Trusted	Minimal oversight
L6	Certified	Audit-only
L7	Sovereign	None

Certification Tiers (8 tiers)

Tier	Name	Score Range	Description
T0	Sandbox	0–199	No external verification
T1	Observed	200–349	Identity registered, monitored
T2	Provisional	350–499	Initial capabilities verified
T3	Monitored	500–649	Continuous monitoring active
T4	Standard	650–799	Standard certification achieved
T5	Trusted	800–875	Full trust established
T6	Certified	876–950	Third-party audit completed
T7	Autonomous	951–1000	Highest assurance level

Three-Layer Architecture

Layer	Name	Purpose
1	Identity & Trust Primitives	WHO (DIDs, OIDC) + WHAT (Domains/Levels/Tiers)
2	Capability Certification & Extensions	What agent can do, how verified
3	Semantic Governance & Runtime Assurance	Behavioral monitoring, drift detection

Relationship to ACI

CAR is the registry and certification layer built on top of the ACI specification. While ACI defines the identifier format and protocol standards, CAR adds:

• Runtime trust scoring (behavioral + certification + context)
• Registry API for agent discovery and management
• Certification workflow and attestation management
• Client SDKs (TypeScript, Python) and CLI tools

Ecosystem

Project	Description	Link
Vorion	First validated CAR implementation	vorion.org
CAR Spec	Categorical Agent Registry standard	npmjs.com/@vorionsys/car-spec
BASIS	Behavioral AI Safety & Integrity Standard	basis.vorion.org
ATSF	Agentic Trust Scoring Framework (46 layers)	atsf.vorion.org
Cognigate	AI governance gateway & policy engine	cognigate.dev
AgentAnchor	Agent identity & attestation registry	agentanchorai.com