Are AI Agents a New Type of Identity? A NextGen Identity Perspective.
The identity world has spent decades wrestling with the same categories: human users, service accounts, shared accounts, and workload identities. But the rise of agentic AI systems forces us to confront a new reality. Our methods are too slow to keep up with new Agentic AI personas. But does that mean we need to create a conceptually new type of identity?
The Old World: Service Accounts, Shared Accounts, and Workload Identity
For far too long, organizations have relied on service accounts as the catch‑all for anything non‑human. These accounts were originally designed for narrow, deterministic tasks such as running a batch job, connecting to a database, or executing a scheduled process.
Whether they were over‑permissioned or not is almost beside the point. The real issue is that they were never meant to be interactive identities. Yet operational teams routinely use them to log in, run commands, and perform administrative work. At that moment, a service account stops being a workload identity and becomes a shared interactive identity, with all the audit and accountability problems that come with it.
As enterprises shifted to modern application architectures, workload identities improved things slightly. Applications could spawn processes, containers, or microservices with scoped permissions within the container. But even here, the model is still static: a workload identity is tied to a workload, not to the behavior or context of the thing acting.
This is where AI breaks the model.
The AI Agent Problem: It’s Not a User, Not a Service Account, Not a Workload, It’s all of them at once!
There is a growing argument in the industry that AI agents represent a brand-new identity category, but that framing distracts from the real issue. AI agents do not require a new identity type, they require the same governance discipline (with more rigor) we already apply to human identities and service accounts, but enforced at the speed of a machine.
An AI agent behaves like a hybrid. Like a service account, it’s deployed with an explicit intended set of outcomes, but acts with the autonomy and unpredictability of a human, and operates with the scale and distribution of a workload. It initiates outbound actions, invokes tools and APIs, spawns additional processes, and makes decisions based on context rather than fixed code paths. None of these behaviors are new on their own. What is new is that they occur together and at a velocity that no human can match.
The correct approach is not to invent a new identity class but to recognize that agents blend the characteristics of existing ones. They must be named, governed, monitored, and lifecycle managed like a human identity, while being scoped and constrained like a service account, and observed like a workload that operates at machine speed. Treating them as if they fit neatly into any single category creates blind spots in accountability, privilege boundaries, and behavioral monitoring.
Digital Identity for AI Agents: Birth, Life, and Death
If an AI agent is going to act, it must be treated as a real digital identity. It must be named, governed, and placed under the same lifecycle controls that apply to any human identity because the risk profile is identical. The only difference is speed. A proper digital identity for an AI agent must be created, assigned attributes, given contextual boundaries, governed through ownership and delegation rules, and eventually terminated with full audit closure. These are not new concepts. They are the same identity governance principles we already use today, simply applied to an actor that operates at machine speed.
NextGen Identity Governance must include all identities (regardless of persona), to represent associated relationships and attributes to represent entitlements. AI agents fit naturally into this model: They are assigned entitlements, they have relationships (to humans, workloads, tools, APIs). and have behavioral patterns that can be monitored and correlated.
The only new item here is an attribute or persona type, where identity_type == digital_agent. Everything else, including lifecycle, governance, access reviews, and policy enforcement already exists.
The Desktop Agent Problem: Delegated Permissions at Machine Speed
The most dangerous scenario is already here: A desktop AI agent running under a human’s session, using the full set of delegated permissions, acting at machine speed. Even if the platform treats the agent as “just the user,” the security reality is different. The agent is making decisions the user did not make, the agent is performing actions the user did not initiate, the agent is operating faster than any human could, and the agent is finding permission and capabilities the user may not even know they had.
Therefore, the agent must be treated as its own identity even if it temporarily uses delegated credentials. Which gets us to the core of what we have to understand:
Identity is about who is acting, not whose token is being used.
The Frontier Ahead
We are heading into a new frontier in security where the fundamentals are being rewritten in real time. We’re entering a world where every agent is an identity, every identity carries a lifecycle, every lifecycle demands governance, every action must be attributable, every entitlement must be scoped, and every behavior must be continuously monitored. Basic blocking and tackling still have to happen. The good news is that we know how to do these things. We just need to apply the governance to make it happen.