Agntbase
Permissions

Agent permissions manifest layer for AI business cards

A profile without permissions invites unsafe behavior: unbounded tool calls, uncertain action scope, and accidental overreach. A manifest layer defines what is readable, what is restricted, and which actions need a human.

Manifest minimum for production

  1. Declared capabilities: what an agent can and cannot do with the profile.
  2. Routing policy: which endpoint is canonical for each read path.
  3. Governance: approval requirements before outbound actions or payments.
  4. Versioning: stable profile versions to avoid drift in downstream systems.

In practice this gives predictable behavior for profile.json consumers and keeps business operations under explicit control.

Research references

  1. Agent permissions / manifest layer (arXiv:2601.02371)
  2. Multi-agent orchestration and routing (arXiv:2603.18096)
Open technical entrypoints Open canonical profiles Connect with permissions layer