The perimeter is identity. The identity is now an agent.

The consensus that the security perimeter is identity, not network, is roughly a decade old. It survived the shift to cloud, the shift to remote work and the shift to SaaS. It is now meeting its first genuinely new adversary, which is the software agent operating with a human's credentials.

The core problem is not novel. Non human principals, meaning service accounts, API keys and machine identities, have outnumbered human ones inside large organisations for years. What is new is that agents combine the persistence of a service account with the discretion of a human user. The existing identity stack is not designed for that combination.

The near term implication for security teams is that agent identity has to be a first class category, not a subtype of user identity. Distinct issuance, distinct scoping, distinct auditing and distinct revocation. Every serious identity vendor is now shipping toward that model. Almost none of them are there yet.

The medium term implication is that the classic who did what audit question splits into two. Which human authorised the action, and which agent executed it. Log formats, SIEM detections and incident response playbooks all need to acknowledge the split.

The security teams that treat this as a routine extension of the identity roadmap will be well positioned. The ones that treat it as a research project will spend the next two years catching up.