AI Agents Inventory and Attributes

AI agents raise governance challenges that do not fit neatly into the standard attribute categories — challenges specific to autonomous, model-driven actors, and moving quickly as the field matures. This chapter names the most consequential of them and, for each, states how a governed AI Agents Inventory addresses it. The point is not that the inventory solves these challenges on its own, but that it is the precondition for addressing any of them — which is why this inventory is positioned as the operational extension of the Enterprise AI Governance Best Practices document, the record its agent-level governance is applied against.

Vendor Coupling and Cross-Platform Fragmentation

The agent-building and orchestration platforms an enterprise uses each maintain their own registry of the agents created within them. Useful as these are, each is coupled to its platform: it covers only that platform’s agents, keyed to that platform’s identifiers, and answers only that platform’s questions. An enterprise running agents across several platforms therefore has not one inventory but many partial ones, none of them enterprise-wide — and anything built outside a platform falls through all of them. The AI Agents Inventory resolves this by being the vendor-neutral system of record that federates over those platform-native registries: it harvests from them as governed sources, decouples its records from any vendor’s identifiers through Semantic Identifiers, and remains the single authoritative record that outlives any individual tool. A registry that lives only inside a vendor’s platform is not an enterprise system of record.

Non-Human Identity Sprawl

Every agent acts under its own non-human identity — API keys, OAuth grants, service accounts — and those identities proliferate faster and with less oversight than human ones. Ungoverned, they accumulate standing access that no one reviews and no one revokes. The inventory ties each agent to the non-human identities and credentials it holds and the access those confer, making the agent’s true reach visible and giving identity and access governance a record to review, scope, and offboard against.

Definitional Churn

What counts as an ‘agent’ is still stabilizing, and the vocabulary shifts — agent, assistant, copilot, workflow, orchestration. An enterprise that waits for the terminology to settle will govern nothing in the meantime. The inventory sidesteps the churn by governing on a stable, behavioral boundary — an actor that takes consequential action or communicates on the enterprise’s behalf without per-output human approval — rather than on any vendor’s label, and by using durable attribute names that describe what an agent does rather than what this year’s products call it.

The Multi-Agent Boundary

Increasingly an ‘agent’ is itself a system of agents that call one another, and it is not always obvious where one governed record ends and another begins. The inventory governs the individual agent as the unit of record and represents a multi-agent system through the relationships among its member agents, so composition is captured as edges in the Enterprise Model rather than collapsed into a single opaque record — the same discipline that distinguishes an application from the system of applications it participates in.

Attribution and Stop Authority

When an autonomous agent acts, two questions must have answers ready before anything goes wrong: who is accountable, and who can stop it. Left implicit, both are discovered only in the middle of an incident. The inventory makes them explicit attributes — a named accountable owner and a named stop authority for every agent — so accountability is assigned in advance and the authority to halt an agent in production is known, not improvised.

Regulated and Consequential Speech

An agent that communicates on the enterprise’s behalf can create binding legal and regulatory consequences through what it says, not only through what it does — a misstatement to a customer, or advice that crosses into regulated territory, can bind or expose the enterprise. Governing this means treating a stakeholder-facing agent the way an enterprise treats a human representative who speaks for it: constrained by policy, trained and attested, and held to the regulations that apply to its domain. The inventory carries the record that makes this defensible — the agent’s served audience, its applicable regulations and obligations, its guardrails, and its attestation that it was trained and authorized to speak as it does.