AI Agents Inventory and Attributes

The AI Agents Inventory governs the enterprise’s AI agents. An AI agent is an autonomous, model-driven software actor that takes consequential action on the enterprise’s behalf — operating on systems and data, and/or producing content that reaches stakeholders by any means, whether a live conversation, a generated email or letter, a document, a notification, or a posted message; electronic or physical; synchronous or asynchronous — in pursuit of a goal, without a human approving each individual action or message, and under its own non-human identity. A single record in this inventory represents one such agent. The defining test is where authority sits: if a human reviews and approves each output before it reaches a system or a stakeholder, the software is a copilot and is governed as a feature of its host application, not here; if the agent’s outputs reach systems or stakeholders without per-output human approval, it is an AI agent and belongs in this inventory.

An AI agent is best understood through its dual nature, and that duality is what makes it its own governed class. It is application-like — designed, deployed, versioned, owned, and operated much like an application or an application-like service — and at the same time person-like, provisioned with an individual identity, a role, access and permissions, training and attestation, and a named accountability and stop authority, much as an organization onboards an individual employee. Standing up an agent is therefore part deploying an application and part onboarding a worker, and the attributes this inventory governs reflect both sides of that nature.

Several adjacent Noun Types are deliberately excluded. An AI agent is not the model it runs on: the model — a large language model or other machine-learning model — is a distinct Noun Type governed in the AI and Machine Learning Models Inventory, and one model can drive many agents. It is not the application it is embedded in or acts upon, which is governed in the Applications Inventory. It is not the tools, APIs, or MCP servers it uses to act, which are reached through Integrations. It is not a bare non-human identity such as a service account or API key; an agent has a non-human identity but is far more than one. And a system of several coordinated agents is not itself a single record — a multi-agent system is understood through the relationships among the individual agent records it comprises.