AI agents as identities

Nova manages AI agents as an identity type of their own, “AI Agent”; in the list of identities they appear on the “AI agents” tab. The same means apply to them as to people – access requests, approvals, recertification and logging – supplemented by rules that apply to agents only.

Mandatory details

An administrator creates an AI agent. Nova requires:

• an accountable person (owner) – an existing identity of the natural person type,
• a purpose describing why the agent exists.

A date for the next review and an expiry date are optional.

Entitlements: capability roles only

AI agents receive entitlements as capability roles with the prefix nova-ai:. For every combination of object (such as identity, business role or account) and action (create, read, update, delete) for which a building block exists, Nova creates such a role – for example nova-ai:identity.update. By default, reading counts as no risk, creating and updating as medium risk and deleting as critical; administrators can change this. The roles take effect only in Nova; Nova does not transfer them to target systems.

Nova keeps these roles separate from all other entitlements:

• When assigning and requesting, Nova allows only capability roles for AI agents; it rejects business roles and other entitlements.
• Only AI agents receive capability roles. They cannot be members of a business role.
• Only the agent’s owner, the owner’s deputy or an administrator can file an access request for an AI agent. An approval workflow can name the “Agent Owner” as approver, see Approval workflows.

What the capability roles govern

Nova can compose steps for lifecycle routines from building blocks with AI. Administrators define which objects and actions AI may use for this under “Administration › AI Assistants” on the “AI Routine Assistants” tab:

• Without an active AI routine assistant with released capabilities, AI composes no steps.
• A routine assistant can be linked to an AI agent. The capability roles the agent validly holds then count as well.
• If the linked agent is inactive, deleted or past its expiry date, the routine assistant no longer counts.
• This only affects composing new steps: saved and built-in steps keep running.

The AI agent itself does not act in Nova. The link makes it the accountable point of reference, not an actor.

Reviews

• Review date: the “Agent Review Scan” job type finds active AI agents whose review date has passed and writes one audit log entry per agent and run – naming the owner and the linked routine assistants. An administrator creates such a job under “Monitoring › Background Jobs” with “New job”; none is preset. The job sends no notifications.
• Recertification: a campaign with the scope “AI Agents” reviews the entitlements of the AI agents; “Agent owner (AI identity)” can be chosen as reviewer. In addition, the matrix of released objects and actions of every linked routine assistant is up for review. If it is revoked, Nova switches the routine assistant off. See Recertification.

When the owner leaves

The preset leaver routine contains the step “Ownership-Übergabe für KI-Agenten” (ownership transfer for AI agents). It transfers the leaving identity’s AI agents to its deputy or, failing that, to its manager according to the organisation. Only an active natural person who is not leaving in the same run can become the successor. If Nova finds none, the agent is left without an owner and Nova writes a warning to the audit log; Nova logs every transfer as well. See Joiners, movers, leavers.

Accounts in Keycloak

In Keycloak, Nova creates a confidential client with a service account for an AI agent instead of a personal account; the client signs in machine-to-machine (client credentials). Its tokens are valid for five minutes by default. “Rotate secret” generates a new client secret; Nova shows it once and does not store it. See Keycloak.