Agentic AI agent identity is raising critical issues for enterprises

Agentic AI is far more complicated than many understand it to be. Delving into the realm of agentic AI, you quickly run across the issue of agent identity. That is, how the identity of an agent is determined, managed, conveyed and operates. Simliar, but very different from humans, the identity of an agent is a critical piece in an agentic AI workflow. To complicate matters, enterprise vendors are taking different paths in how they address agent identity and why.

What is agent identity?

First off, agent identity is just as it sounds…the identity of an agent or how we identify specific agents. This is critical in understanding who…or what is performing the work. In terms of observability and management, we need a clear way to understand and track agents.

In multi-agentic workflows, agents are strung together to perform a task or complete work. Each agent needs its own identity to differentiate between agents.

Should agents be treated like humans?

Since the start of agentic AI, there is an outstanding question of whether AI agents should be treated like how we track and treat humans. The short answer is no.

While there are some traits one could argue are like humans (policies, guardrails, access), agents work in very different ways from humans. One example is that agents can work autonomously from humans on multiple levels. At the same time, agents still need specific guidance on what actions to take as they lack the cognitive capabilities. More on that aspect in a future discussion.

Agents also work at a speed unlike humans. Humans take time to think and act. Agents are much quicker in comparison and therefore can go sideways and further down a bad path before anyone realizes.

Unlike creature comforts and support for humans, agents do not have feelings nor need to navigate human interactions in the same way. If an agent is misbehaving, it is straightforward to take corrective action or simply terminate the agent.

Agents are far more dynamic than humans. Humans do have learning capabilities but cannot learn at the scale and speed of agents.

These are just a few of the many reasons that require us to treat agents differently than humans.

Agent identity dynamics

Moving into identity dynamics, there are several aspects around agent identity that create further complexity to how we should think about agent identity.

• Identity: The identity of a specific agent.
• Agency: Is the agent working on behalf of a person or another agent?
• Authority: Who gave the agent authority and what authority does it have?
• Permissions: What is the agent allowed to and not allowed to do? This is often where policies, guardrails and governance come into play.

Agent lifespan and scope

The lifespan of an agent can range from forever to ephemeral and any number of ways in between. The lifespan and scope of an agent play a role in an agent’s identity. Let’s look at some examples.

• Independent: These agents are delegation vehicles that work independently. These may also be autonomous agents that are dedicated or shared between other users and/or agents.
• Human in the loop: These agents are capable of performing certain tasks but may require a human to complete the full body of work. Because of the human in the loop, it changes the dynamics of what the agent may or may not be able to do independently.
• Ephemeral: I wrote about ephemeral agents last year. Ephemeral agents are dynamically created in real-time as the need arises. Depending on how they are created and the governance around how they are created often dictates the capabilities of ephemeral agents.

Each of these demonstrate varying degrees of independence for the agent. The identity of an agent is closely tied to the lifespan and scope of an agent.

Registration of agents

Many vendors are using an agent registry as a single repository to centrally manage agents. A central repository offers several valuable tools for managing agents from control and governance to awareness and observability.

Agent registries require a bit of organizational control and discipline to leverage effectively. In the advent of personal agents, it is possible for rogue or semi-rogue agents to run outside of the registry. I recently wrote about the risks of personal agents as a potential trojan horse.

Agent registries can also be limiting in certain, more complex architectures. As agents become more sophisticated with capabilities around discovery, learning and reasoning, how registries operate will need to become more dynamic.

Should agents be connected to humans?

One aspect that is drawing sides is if agents are required to be tied to a human or not. As an example, Microsoft has taken the stance that all agents must be assigned to a human while most other vendors are not making this a requirement. There are pros and cons to each approach.

Tying an agent to a human creates provenance of who created the agent and is (potentially) running the agent. The downside is that it creates a hard connection to a human that may or may not be the one ultimately leveraging the agent. If the human changes roles or leaves the org, then reassignment of agents must take place.

Taking a more fluid approach of not requiring agents to be assigned to a human creates autonomy for agents but also could bring orphaned agents into view. Again, discipline and governance are required.

Related to the human connection decision is who pays for the resources when an agent runs. Individual agents may be tied to a user’s cost center. Autonomous agents or those that are shared create a different challenge at runtime. Add in the complexity of multi-step agentic workflows, ephemeral agents and one can see the complexity.

CIO Perspective

Unlike much of the public conversation around agent identity suggesting it is a simple task, agent identity is far more complicated than simply assigning identity and registering agents. For simpler agentic AI use cases, the process is straightforward. However, as workflows mature and become more sophisticated, so will the need to evolve how we think about agent identity.

An agent, is not an agent, is not an agent. Agents are not singular ‘things’ and are as varied as colors in a spectrum. From super agents to sub agents, the role, identity and capabilities of agents will vary greatly. As learning and reasoning become more commonplace, it creates the need for a more dynamic environment in which to operate and manage agents.

Beyond the borders of their own architectures, CIOs will need to account for agents that are engaging MCP and A2A methodologies. How is identity handled by these third parties? Governance will need to play a strong role while maintaining flexibility so as not to kill innovation.

This post is only talking about agent identity. However, ancillary aspects play a role such as resource consumption, compliance, observability, security and governance. Each of these has their own set of considerations and plays a role in both defining an agent’s identity but also providing guidance on how the identity evolves over time.