CrowdStrike introduces AI agent identities: who decides what an agent may do?

CrowdStrike introduced Agentic Identity Provider, or Agentic IdP, at Fal.Con 2026 on September 2. The company says it registers AI agents, gives them verifiable identities, and connects their actions to the human or system they represent. The announcement highlights a fundamental developer question: when software acts for a person, how do we establish who did what?

Identity is not the access decision

The company's explanation separates Agentic IdP's role in establishing an agent's identity from Continuous Identity's role in deciding access. It also describes short-lived, tightly scoped access tokens. That is the vendor's account of its design, not a conclusion from an independent test we performed.

Consider a team designing an agent to help customer support. “Which agent is this?” differs from “may it only draft a reply, or may it send the message?” The team should make both decisions deliberately rather than treating identification as automatic permission to carry out every action.

The answer can vary with the task. A draft awaiting review is different from a message sent to a customer, just as reading a record is different from changing it. These are design examples to explain the distinction, not a list of capabilities we tested in the announced product.

Following the chain of responsibility

In our assessment, a useful evaluation example is following a task from a person's request to its outcome. What request started the work? Which steps were performed? Where did a person make a decision? Those answers explain the workflow rather than merely displaying the final result.

If every action appears under one employee's name, interpreting an error can become harder. This is an organizational and technical reason to distinguish the human's role from the software's role. It is not a report about an incident we observed at a particular organization.

An announcement is not proof that every feature has shipped

CrowdStrike's release cautions that some unreleased services or features remain in development and are subject to change. A team therefore needs to check what is actually available before making an adoption decision, rather than assuming that every part of an announced vision is immediately usable.

Our suggested starting point is a bounded task. Define what the agent may read, what it may write, and what requires a person's approval. Then review how the task can be stopped and who will handle an incorrect result.

A clear identity provides an important starting point. The limits on action still require a separate decision, connecting the software's permissions to the job it was created to perform and the person accountable for the outcome. That distinction matters even when the software can act quickly and the interface makes delegation look simple.

Sources

• CrowdStrike — September 2 announcement
• CrowdStrike — Agentic Identity Provider explanation