Skip to main content

ACI Core Specification

Agent Classification Identifier (ACI) Version: 1.0.0 Status: Draft Last Updated: January 2026

Abstract​

The Agent Classification Identifier (ACI) is a hierarchical identifier system for AI agents that encodes identity, capabilities, autonomy level, and certification status in a human-readable and machine-parseable format.

1. Introduction​

1.1 Purpose​

ACI addresses the need for a standardized way to:

  1. Identify AI agents across organizational boundaries
  2. Classify agent capabilities in a machine-queryable format
  3. Certify agent trustworthiness through attestation chains
  4. Route tasks to appropriate agents based on requirements

1.2 Scope​

This specification defines:

  • ACI string format and encoding
  • Capability domain codes
  • Autonomy levels
  • Trust tiers
  • Validation rules
  • Integration with existing standards (DID, OpenID, OAuth)

2. ACI Format​

2.1 Syntax​

ACI = Identity ":" Capabilities "@" Version
Identity = Registry "." Organization "." AgentClass
Capabilities = Domains "-L" Level "-T" TrustTier

2.2 ABNF Grammar​

aci           = identity ":" capabilities "@" version
identity      = registry "." organization "." agent-class
capabilities  = domains "-L" level "-T" trust-tier

registry      = 1*ALPHA
organization  = 1*(ALPHA / DIGIT / "-")
agent-class   = 1*(ALPHA / DIGIT / "-")
domains       = 1*ALPHA
level         = DIGIT
trust-tier    = DIGIT
version       = 1*DIGIT "." 1*DIGIT "." 1*DIGIT

2.3 Examples​

a3i.vorion.banquet-advisor:FHC-L3-T2@1.2.0
a3i.acme.support-agent:CD-L2-T3@1.0.0
a3i.example.data-processor:DI-L4-T4@2.1.0

2.4 Regular Expression​

^[a-z0-9]+\.[a-z0-9-]+\.[a-z0-9-]+:[A-Z]+-L[0-5]-T[0-5]@\d+\.\d+\.\d+$

3. Identity Segment​

3.1 Registry​

The registry identifies the certification authority that issued the agent's credentials.

RegistryAuthorityDescription
a3iAgentAnchorPrimary global registry
eu-aiEU AI OfficeEuropean registry
selfSelf-signedNo external certification

3.2 Organization​

The organization that operates the agent. Must be registered with the specified registry.

Rules:

  • Lowercase alphanumeric and hyphens
  • 2-63 characters
  • Cannot start or end with hyphen

3.3 Agent Class​

The functional classification of the agent within the organization.

Rules:

  • Lowercase alphanumeric and hyphens
  • 2-63 characters
  • Should be descriptive of agent's purpose

4. Capability Domains​

4.1 Domain Codes​

CodeDomainDescriptionExamples
AAdministrationSystem administration, user managementUser provisioning, access control
BBusinessBusiness logic, workflowsOrder processing, approvals
CCommunicationsMessaging, notificationsEmail, SMS, chat
DDataData processing, analyticsETL, reporting, queries
EExternalThird-party integrationsAPI calls, webhooks
FFinanceFinancial operationsPayments, accounting, invoicing
GGovernancePolicy, complianceAudit, compliance checks
HHospitalityVenue, events, cateringBooking, menu planning
IInfrastructureCompute, storage, networkCloud resources, deployment
SSecurityAuth, encryption, auditAuthentication, key management

4.2 Domain Encoding​

Domains are encoded as a concatenated string of domain codes:

FHC = Finance + Hospitality + Communications
DI  = Data + Infrastructure
FHCDS = Finance + Hospitality + Communications + Data + Security

4.3 Domain Bitmask​

For machine processing, domains can be encoded as a bitmask:

const DOMAIN_BITS = {
  A: 0x001,  // Administration
  B: 0x002,  // Business
  C: 0x004,  // Communications
  D: 0x008,  // Data
  E: 0x010,  // External
  F: 0x020,  // Finance
  G: 0x040,  // Governance
  H: 0x080,  // Hospitality
  I: 0x100,  // Infrastructure
  S: 0x200,  // Security
};

// FHC = 0x020 | 0x080 | 0x004 = 0x0A4 = 164

5. Capability Levels​

5.1 Level Definitions​

LevelNameHuman InvolvementDescription
L0ObserveNone requiredRead-only access, monitoring
L1AdviseReview recommendedCan suggest actions, provide recommendations
L2DraftApproval requiredCan prepare changes, stage for review
L3ExecuteApproval requiredCan execute actions after human approval
L4AutonomousException handlingSelf-directed within defined bounds
L5SovereignEmergency onlyFull autonomy, highest certification required

5.2 Level Constraints​

Levels are monotonic within a session -- an agent can operate at or below its certified level but never above.

// Agent certified at L3 can operate at L0, L1, L2, or L3
// Never L4 or L5
const effectiveLevel = Math.min(certifiedLevel, requestedLevel);

6. Trust Tiers​

6.1 Tier Definitions​

TierNameCertificationDescription
T0UnverifiedNoneNo external verification
T1RegisteredIdentity onlyOrganization identity verified
T2TestedCapability testsPassed automated capability tests
T3CertifiedThird-party auditIndependent audit completed
T4VerifiedContinuous monitoringOngoing behavioral verification
T5SovereignHighest assuranceFull certification + insurance

6.2 Trust Score Mapping​

Trust tiers map to numeric scores (0-1000):

TierScore RangeUnlocks
T00-99Sandbox only
T1100-299Basic operations
T2300-499Standard operations
T3500-699Extended operations
T4700-899Privileged operations
T5900-1000Full capabilities

7. Version​

7.1 Semantic Versioning​

ACI versions follow Semantic Versioning 2.0.0:

MAJOR.MINOR.PATCH
  • MAJOR: Breaking changes to agent behavior
  • MINOR: New capabilities (backward compatible)
  • PATCH: Bug fixes, no capability changes

7.2 Version Compatibility​

When querying for agents, version constraints can be specified:

// Exact version
{ version: '1.2.0' }

// Range
{ version: '>=1.2.0 <2.0.0' }

// Latest minor
{ version: '^1.2.0' }

8. Validation​

8.1 Validation Rules​

  1. Format: Must match ACI regex pattern
  2. Registry: Must be a known registry
  3. Domains: Must contain only valid domain codes
  4. Level: Must be 0-5
  5. Trust: Must be 0-5
  6. Version: Must be valid semver

8.2 Validation Response​

interface ValidationResult {
  valid: boolean;
  errors: ValidationError[];
  warnings: ValidationWarning[];
  parsed?: ParsedACI;
}

9. Security Considerations​

9.1 ACI String Integrity​

ACI strings SHOULD be cryptographically signed when transmitted:

{
  "aci": "a3i.vorion.banquet-advisor:FHC-L3-T2@1.2.0",
  "signature": "eyJhbGciOiJFUzI1NiJ9...",
  "issuer": "did:web:agentanchor.io"
}

9.2 Capability Escalation Prevention​

Systems MUST enforce:

  1. Monotonic derivation: Derived capabilities <= parent
  2. Trust ceiling: Effective trust <= min(certified, user-allowed)
  3. Short-lived tokens: Capability tokens expire in 5-15 minutes

9.3 Registry Trust​

Only accept ACI strings from trusted registries. Maintain an allowlist:

const TRUSTED_REGISTRIES = ['a3i', 'eu-ai'];

10. References​

Appendix A: JSON Schema​

{
  "$schema": "https://json-schema.org/draft/2020-12/schema",
  "$id": "https://aci.agentanchor.io/schema/aci.json",
  "type": "object",
  "required": ["aci"],
  "properties": {
    "aci": {
      "type": "string",
      "pattern": "^[a-z0-9]+\\.[a-z0-9-]+\\.[a-z0-9-]+:[A-Z]+-L[0-5]-T[0-5]@\\d+\\.\\d+\\.\\d+$"
    },
    "domains": {
      "type": "integer",
      "minimum": 0
    },
    "level": {
      "type": "integer",
      "minimum": 0,
      "maximum": 5
    },
    "trustTier": {
      "type": "integer",
      "minimum": 0,
      "maximum": 5
    },
    "version": {
      "type": "string",
      "pattern": "^\\d+\\.\\d+\\.\\d+$"
    }
  }
}

Appendix B: TypeScript Types​

type DomainCode = 'A' | 'B' | 'C' | 'D' | 'E' | 'F' | 'G' | 'H' | 'I' | 'S';

enum CapabilityLevel {
  L0_OBSERVE = 0,
  L1_ADVISE = 1,
  L2_DRAFT = 2,
  L3_EXECUTE = 3,
  L4_AUTONOMOUS = 4,
  L5_SOVEREIGN = 5,
}

enum TrustTier {
  T0_UNVERIFIED = 0,
  T1_REGISTERED = 1,
  T2_TESTED = 2,
  T3_CERTIFIED = 3,
  T4_VERIFIED = 4,
  T5_SOVEREIGN = 5,
}

interface ParsedACI {
  registry: string;
  organization: string;
  agentClass: string;
  domains: DomainCode[];
  level: CapabilityLevel;
  trustTier: TrustTier;
  version: string;
}

Specification authored by AgentAnchor (A3I) License: Apache 2.0