AI Agent Catalog

The AI agent catalog is a mandatory primitive for any guardian-agent deployment per Gartner. It inventories all AI agents — registered, unregistered, official, custom, third-party, shadow, or rogue — within an organization’s network. It scores risks and tracks them over time. It stores agent cards as metadata.

The catalog is the foundation everything else builds on: you cannot govern, monitor, or enforce policy on agents you cannot enumerate.

Two roles for the catalog

The catalog plays a dual role in the wiki — independently arrived at by two Gartner publications:

LensSourceThe catalog is…
Security inventory primitiveMarket Guide for Guardian Agents (Feb 2026)The mandatory enumeration substrate for guardian agents — visibility, risk scoring, runtime policy attribution
Procurement coordination primitiveScaling Agentic AI talk (May 2026)The single source of truth that procurement uses to vet new agent purchases against the existing stack, prevent duplication, and insert IT requirements at “zero day” of any new RFP/RFI

The two roles share the same artifact (the agent card) but use it differently. The procurement role makes the catalog the chokepoint where new agentic services enter the enterprise — a posture that complements the runtime-enforcement role from the Market Guide. Both roles are required for the AI Agent Layered Council to function: Procurement uses the catalog to coordinate purchases, the guardian agent / oversight layer uses it to enforce policy at runtime.

What the catalog must contain

Per Gartner’s mandatory feature definition:

Field classExamples
IdentityUnique agent ID; cryptographic identity (SPIFFE SVID, Okta agent ID, Microsoft Entra Agent ID); publisher signature
CapabilitiesWhat tools the agent can call; what data it can access; what autonomy tier it operates at
Interaction endpointsAPIs, gateways, MCP servers it consumes or exposes
Authentication requirementsWhat credentials, scopes, or tokens it needs to operate
LineageWho created it, when, from what template; deployment history
Risk scoreComputed from capabilities × data access × autonomy × usage history
Owner mappingHuman owner (responsible party) + machine owner (parent agent or platform)
StatusActive, deprecated, sandboxed, blocked, decommissioned

This metadata bundle is what Gartner calls an agent card — analogous to a SaaS app’s profile in a CASB inventory, but for agents.

Discovery: registered + unregistered + shadow + rogue

The catalog must enumerate all four populations:

PopulationHow they’re discovered
RegisteredThe agent self-registers with the IAM / agent platform on creation
UnregisteredDiscovered via network telemetry, identity provider observation, or platform-API enumeration; backfilled into the catalog
ShadowShadow AI / shadow automation — agents created outside sanctioned platforms (developer-side, BYOAI, ungoverned IDE extensions)
RogueAgents whose behavior diverges from declared intent or whose identity has been compromised

No catalog discipline = no governance. This is the entry-level failure mode for AI agent security programs.

Risk scoring

The catalog must score risk per agent and track over time. Inputs:

  • Capability surface (tool count, sensitivity of accessible APIs)
  • Data access scope (which classification levels, which sources)
  • Autonomy tier (per Least Agency Principle: auto / notify / confirm / block)
  • Usage frequency and patterns
  • Behavioral baseline drift (agent behavioral monitoring signal)
  • Recent incidents involving this agent or its dependencies (skill registry, MCP server, model)
  • Owner status (active human owner vs. orphaned)

Risk scores feed back into runtime decisions: high-risk agents get tighter Operative oversight; low-risk agents get a lighter touch.

Implementations (2026)

Vendor / ToolCoverage
Microsoft Agent 365First-party + third-party agents registered with Entra Agent ID
Okta for AI AgentsCross-platform agent identity registry (GA April 30, 2026)
Astrix Security, AembitNon-human identity discovery + governance
CrowdStrike, SentinelOneEndpoint-level agent discovery via process telemetry
Wiz AI-SPM, Palo Alto Prisma AIRSAI asset inventory across cloud environments
Internal CMDB extensionsMany enterprises extend their CMDB to track AI agents

No single tool covers all four populations comprehensively. Gartner’s prediction is that independent guardian-agent vendors will provide unified catalog discovery as a category in 2027–2028, displacing the current hyperscaler-specific tools for cross-vendor enterprises.

Where this fits the wiki

Wiki pageConnection
Non-Human Identity (NHI)Catalog is the inventory layer for NHI; agent cards = NHI metadata
AI Agent Identity ArchitectureThe architectural reference for how catalog identities are assigned and used
Guardian AgentCatalog is mandatory feature category 1 (visibility and traceability)
Agentic AI Security Capability Maturity Model — A 2026 Practical Proposal D2 IdentityCatalog L3+ requirement: comprehensive across all four populations
Shadow AI, Shadow AutomationCatalog discovery surfaces both

CMM L3+ evidence requirements

For an organization to claim Level 3 on D2 Identity & Authorization in the CMM, the catalog evidence must include:

  1. Comprehensive agent inventory with documented coverage of all four populations (not just registered)
  2. Agent cards for every cataloged agent with the field set above
  3. Risk-score methodology documented and applied uniformly
  4. Catalog refresh cadence (continuous discovery, not point-in-time)
  5. Owner attribution for ≥95% of cataloged agents (orphan detection for the rest)

L4 adds: integration with runtime decisions (catalog signals feed Operative behavior); behavioral baselines per agent.

L5 adds: cross-vendor federation (catalog spans Microsoft + AWS + Google + on-prem with unified ID).

Open issues

Catalog gaps

  1. Cross-vendor agent identity reconciliation. No standard yet for resolving “is the Microsoft Entra Agent ID for agent-X the same agent as the Okta agent ID for agent-X?” Identity federation across vendors is unsolved.
  2. Shadow agent fingerprinting. Without declared identity, fingerprinting depends on behavioral metadata (model used, tools called, output style). Gartner predicts metadata fingerprinting becomes the fallback identity in 2026–2027.
  3. Skill / MCP-server inventory. Catalog of skills and MCP servers the agents consume is a separate (but related) inventory problem. See MCP Security.

See Also