Agent access should look like delegated work¶
AI agents should not appear in a system as anonymous API traffic or as another copy of a user's long-lived credential. They should appear as delegated actors with explicit scope, expiry, and approval context.
That framing changes the shape of agent authentication. The important question is not whether an agent can hold a token. The important question is whether the tenant can understand why that token exists, which user or workflow authorized it, which tools it can reach, and when the delegation ends.
SigID models agent access around constrained delegation. Human identity remains the anchor, but the agent receives a bounded credential that can be audited and revoked without resetting the user's account. For MCP servers and tool APIs, that creates a cleaner contract than reusing user sessions or distributing static service credentials.
Good agent access has a few visible properties:
- Every agent credential names its purpose.
- Tool access is scoped narrower than the user's full account.
- Approval and consent are recorded before execution.
- Expiry is short enough to limit unattended access.
- Revocation does not require rotating unrelated application secrets.
This is why agent identity belongs beside OAuth, consent, organizations, and audit logs in the product model. The system should make delegated work legible.