What is an agent passport?
An agent passport is the identity document of a single AI agent: who owns it, what it does, which model it uses, what it is allowed to do, at which autonomy level it works and how it has been tested. Without a passport, an agent does not go into production.
- ID
- ENR-OPS-001
- Status
- Pilot
- Level
- 3
| Field | Example: outage reports agent at an energy company |
|---|---|
| Owner | Head of customer service |
| Purpose | Diagnose failed charging sessions and resolve or escalate them |
| Trigger | New report via phone, app or WhatsApp |
| Models and provenance | Sprinter for intake, workhorse for diagnosis; version pinned, provenance recorded |
| Hosting | Own cloud environment in the EU |
| Systems and permissions | Charging platform read, ticket system create and update |
| Data sources | Session logs, telemetry, contract terms, knowledge base |
| Human in the loop | Credit notes above the threshold and every escalation |
| Autonomy level | Level 3: acts, a person approves |
| EU AI Act risk class | Limited risk, assessed by Legal |
| Evaluations | Golden set of real reports, prompt injection tests |
| Last review | At go-live, then every quarter |
Explanation
The passport is stored next to the agent's configuration and reviewed every quarter. Anyone who wants to extend an agent updates the passport first.
For open-weight models we add provenance fields: which lab, which version, which licence and where the model runs. A model's origin has become a security question.
Software vendors are now building their own variants: ServiceNow announced AI Passport and Workday has an Agent Passport. The passport as we describe it works on any platform, including when agents run on several platforms.
