AI Agents Inventory and Attributes

Section A — Sourcing and Harvesting

Before building the AI Agents Inventory from scratch, assess how much of it can be harvested from systems already operating in the enterprise. AI agents almost always exist — at least partially — in other platforms before a formal inventory is established. The most valuable sources are the platform-native registries maintained by the agent-building and orchestration platforms the enterprise already uses, which list the agents created within each; the identity and access systems that hold the non-human identities and credentials agents authenticate with; the cloud and API gateways that see the calls agents make; and the integration and MCP-server catalogs that reveal what agents connect to. Harvesting from these sources reduces the initial data-entry burden, accelerates time to Crawl completeness, and — critically for agents — surfaces records that manual discovery would miss.

That last point is the defining challenge of this inventory. The highest-risk agents are precisely the ones no one has registered — stood up by a team for a project, auto-provisioned by a SaaS platform, or wired together through an integration without central oversight. These shadow agents cannot be governed, secured, or defended, because no one knows they exist. Building the inventory is therefore as much a discovery exercise as a data-entry one: reconcile the harvested platform-native registries against identity systems, gateway logs, and integration catalogs to find the agents that appear in one source but not the others, and treat every unregistered non-human identity that acts autonomously as a candidate agent to triage.

AI agents can themselves assist the harvesting — for example, by reconciling records across platform registries and identity systems, or by drafting initial records from documented sources — but AI-generated records are a starting point requiring human validation, not authoritative records, and their generation method should be documented in the Provenance and Audit Attributes category so their origin is transparent and auditable. Where harvesting is not possible, the inventory must be built and maintained manually; this is entirely viable at Crawl and Walk maturity and is not a failure mode. The most important thing is that the inventory exists, is accurate, and is actively maintained — the automation can come later.

Section B — Ownership and Accountability

Every inventory must have a named owner — an individual or function that is accountable for the accuracy, completeness, and governance of the inventory as a whole. Inventory ownership is distinct from the ownership of individual AI Agent records: a record owner is accountable for the attributes of one specific AI Agent; the inventory owner is accountable for the schema, the governance process, and the overall health of the AI Agents Inventory as a governance artifact.

For inventories with a natural organizational home — where a specific function already governs or consumes the AI Agents — ownership should be assigned to that function. For inventories with no clear organizational owner, the IF4IT recommends assigning ownership to a cross-functional function such as Enterprise Architecture, which already governs the Enterprise Model of which this inventory is a component, or Software Engineering, where the AI Agents are primarily technology-facing. Ownership by committee without a named accountable individual is not recommended — it produces diffused accountability and inconsistent governance.

Section C — Lifecycle and Review Cadence

The AI Agents Inventory is a living governance artifact. It must be actively maintained through a formal lifecycle — not treated as a one-time deliverable. Every AI Agent record moves through defined lifecycle states: Proposed, Active, Under Review, Deprecated, Retired, and Archived as appropriate for this Noun Type. Lifecycle state is governed by the inventory owner and documented in the Lifecycle and Status Attributes category of each record.

Reconciliation cadence — how often the inventory is reviewed and reconciled against source systems or organizational reality — should be established and enforced: at Crawl maturity, quarterly reconciliation at minimum; at Walk maturity, monthly reconciliation, or event-driven on significant organizational or technology changes; at Run maturity, continuous or near-continuous, automated where possible, with human review for exceptions.

Reconciliation that happens informally and undocumented is reconciliation that did not happen for governance purposes. Establish a formal reconciliation event with documented outcomes, owned by the inventory steward.

Section D — Data Quality and Starting Approach

Do not attempt to populate all attributes for all AI Agents simultaneously. The most common failure mode for inventory initiatives is scope overreach — producing a large volume of incomplete records and losing organizational confidence in the data before any governance value is delivered.

Recommended approach: (1) Identify all known AI Agents and create a stub record for each — Semantic ID, Display Name, and Description only — producing a complete, if thin, inventory. (2) Populate all remaining Crawl attributes across all records before any Walk attributes are added to any record. (3) Validate Crawl completeness — 100% of known AI Agents with 100% of Crawl attributes populated — before advancing. (4) Populate Walk attributes systematically, one category at a time if necessary. (5) Introduce Run attributes only when tooling and pipeline maturity justify automated derivation and calculation.

Quality thresholds: Crawl attributes must be 100% complete before Walk begins. All records should be validated against at least one authoritative source at creation and at each reconciliation. Records that cannot be validated should be flagged with a data quality indicator in the Provenance and Audit Attributes category.

Section E — Access Control

The AI Agents Inventory contains governance-sensitive data. Access should be governed explicitly: read access broadly available to consuming stakeholders; write access restricted to the inventory steward, designated data owners, and authorized automated feeds; schema change access reserved for the inventory owner and governing body. At Crawl maturity, a shared spreadsheet with restricted edit access is a fully adequate access control model.

Section F — Change Management

The attribute schema of the AI Agents Inventory is itself a governed artifact. Schema changes — adding, removing, or renaming attributes — follow a five-step process: Propose, Review, Approve, Implement, and Communicate. Schema changes should not be made during an active reconciliation cycle. Changes must be recorded in the inventory document’s change log.

Section G — Archival and Retention

When an AI Agent is retired, its record is not deleted — deletion destroys audit history. Update the Lifecycle Status to Retired, retain the record for one full reconciliation cycle in the active inventory, then archive it. Archived records remain queryable but are excluded from active governance reporting. Retain indefinitely any record for an AI Agent involved in a significant decision — acquisition, major investment, compliance finding. For all others, define a retention period consistent with applicable regulatory requirements and organizational policy.