AI Agents
The AI Agents module (sidebar, "AI Security" group) is the register and governance record of the organisation's AI agents: who owns each one, which model and vendor it uses, what data it accesses, what tools it has, how autonomously it acts, which credentials it uses (metadata only) and when it should be reviewed. Each agent is linked to the model card, to the EU AI Act (section 36), to AI incidents and to ARIA assessments (section 34).
Who can do what: SUPER_ADMIN and ADMIN can do everything; the MANAGER can read, register and update; AUDITOR, VIEWER and USER can only view (they do not see the write buttons and cannot open the new agent form).
Agent list
For each agent, the list shows the Owner, Risk, Autonomy (Suggests only / Acts with approval / Acts alone), Completeness, Next review and pending Drafts. An agent with no owner appears as No owner. The next review is calculated from the risk (6 months for High or Unacceptable risk, 12 months for the rest) and also appears in the calendar.
Registering an agent by hand
- Click New agent
- Fill in the form, by section: Identification and accountability (Name, Purpose, Description, Owner, Entity, Environment, Risk level, Status), Model and vendor, Data and privacy, Tools and autonomy, Credentials and access and Human oversight and logs. Only the Name is required to save; the rest can be completed later
- Under Risk level, leave Automatic (suggested) to use the risk calculated by the assessment
- Under Tools, use Add tool to enter the name, description, the impact (low, medium or high) and whether it is With approval from a human
- Under Credentials (metadata only), use Add credential to enter the type (API key, OAuth token, service account…), the Scope (what it allows), Who manages it and the Rotation. Do not record the values of keys, tokens or passwords: the form rejects text that looks like a secret
- Click Save. You are taken to the agent's record
Registering an agent with Vera
- In the list, click Register with Vera (on an existing agent's record, Update with Vera). The assistant opens with the request already written
- Vera asks one question at a time, only about what is still missing. Never ask her for, or give her, values of keys or passwords: only the type and scope
- Vera does not register the agent: she creates a draft and shows an AI agent draft card with the Completeness, the fields filled in and What is missing
- On the card, you decide:
- Confirm: registers the agent (Agent registered.) and lets you Open the record
- Edit: opens the pre-filled form to correct before approving
- Reject: discards the draft (Draft rejected.)
- Confirmation is always done on the card, never through a chat message
Drafts awaiting confirmation
Pending drafts (from Vera or from a form) appear at the top of the list, under Drafts awaiting confirmation, with the origin (Vera or Manual), who created them, how many fields are missing and when they expire (7 days). Nothing is registered until a person clicks Approve (or Review and approve, which opens the form with Save and approve) or Reject. The person who approves is recorded as the reviewer.
The agent record
The record brings together the profile sections, the risk, the completeness bar and:
- What is missing and why: the fields still to be filled in, grouped by theme, each with the reason and the framework that requires it (ISO/IEC 42001, AI Act, GDPR, OWASP for agents). When everything is filled in it shows The record is complete.
- Governance alerts: DPIA suggested (the agent processes personal data and acts above "suggests only"), high-impact tools that execute without human approval, duty of transparency when it interacts with people (AI Act Art. 50) and the Suggested risk when it differs from the recorded one (the decision is yours)
- Credentials: only type, scope, manager and rotation; the values are never stored
- Links: Model card, EU AI Act, AI Incidents and Aria Assessments
Generating tasks
On the record, Generate tasks creates one task per theme with missing fields (assigned to the owner, who is notified) and, if there is a DPIA alert, a task "Assess the need for a DPIA". The DPIA is not created automatically. The message states how many tasks were created and how many already existed; tasks that are still open are not duplicated.
Preparing the Aria assessment
Prepare Aria assessment creates the ARIA assessment linked to the agent and pre-fills the context (tools, autonomy, data and access) from the record (Aria assessment created and pre-filled.). If one already exists, it is reused (Existing Aria assessment reused.).
AI Register
The AI Register shows the organisation's AI systems and the type of each (model, system or agent); EterShield's own AI capabilities no longer appear mixed in. Agents open on the record above.