Definition​

An agent passport is a portable identity and trust document for an AI agent. It binds together an agent identity, a human or organizational principal, allowed capabilities, signed intents, claims, expiry, and revocation data.

The problem​

Most systems today can identify a human account or a service credential. They usually cannot answer:

• Which agent is acting?
• Was this specific action approved by the user?
• What capabilities were delegated?
• Which claims about the agent are signed by someone trustworthy?
• How can access be revoked across systems?

A practical model​

A useful agent passport should be boring and implementation-friendly:

1. stable agent identifier;
2. owner/principal identifier;
3. scoped capabilities;
4. signed intents for high-risk actions;
5. short expiry windows;
6. revocation endpoint or registry;
7. audit references.

Why signed intents matter​

A signed intent turns “the agent said the user wanted this” into a verifiable statement:

{
  "subject": "user:yaroslav",
  "agent": "agent:writing-assistant",
  "action": "publish_article",
  "scope": ["domain:blog.popov.app"],
  "expires_at": "2026-07-31T23:59:59Z"
}

The receiving system can verify the signature, scope, expiry, and revocation state before accepting the action.

Open questions​

• Should agent passports be closer to DIDs, OAuth clients, certificates, or package manifests?
• How should claims be distributed and revoked?
• Which actions require explicit signed intents instead of ambient capability grants?
• How can this work for local-first agents without a centralized identity provider?

Related notes​

• Agent passports
• Signed intents