Identity · CA
Cisco's Duo Now Treats Agents as Directory-Native Identities
Cisco's push into agent security, detailed this week by VentureBeat, reveals a specific architectural choice worth noting: Duo isn't just adding an agent category to its identity platform — it's making agents first-class citizens of the directory itself, alongside humans and service accounts.
The mechanism matters more than the branding. For non-human identities, Duo delegates scoped permissions through OAuth, using the same authorization specifications that MCP already relies on. That's a deliberate alignment choice: rather than building a parallel permissioning scheme for agents, Cisco is wiring into the token-scoping model the Model Context Protocol ecosystem has already settled on. It also reduces the risk of credentials being passed inadvertently to an agent or deliberately to an attacker, since scoped OAuth tokens narrow what an agent can do with a given grant rather than handing over a standing credential.
The Astrix acquisition slots in as the discovery layer underneath that runtime identity model: Cisco's purchase extends Duo's work into discovery, showing customers the non-human identities in their environment and the accounts, permissions, and secrets those identities use. In other words, the stack splits into two problems that have to be solved together — finding every agent that already exists (Astrix) and then giving each one a scoped, directory-tracked identity it has to use going forward (Duo).
What's notable isn't that a vendor is bolting agent support onto existing IAM — everyone is doing that right now. It's the specific decision to anchor agent authorization to MCP's existing OAuth conventions rather than inventing a Cisco-specific scheme. If that holds, it's a small but real signal that MCP's authorization layer is becoming the de facto interchange format that enterprise identity platforms build against, the way SAML and OIDC became the interchange format for human SSO. A vendor the size of Cisco choosing to plug into an existing agent-native spec, instead of forcing agents into a scheme designed for people, suggests the industry may avoid the worst version of agent identity fragmentation — at least at the authorization layer, even as discovery and runtime monitoring remain vendor-specific.