Agent Permissions

Agentic Fabriq Agent Permissions is a permission enforcement system designed for AI agents that act on behalf of individual users. It addresses the problem of shared service accounts granting uniform access to all users regardless of their actual entitlements, meaning an agent acting for a low-privilege user could inadvertently exercise the same access as a high-privilege user. The system works by resolving permissions at the tool-call boundary — outside the model itself — rather than relying on prompt-level instructions. For every request, the acting user's identity is carried along, and the agent's effective permissions are computed as the intersection (overlap) of what the agent has been granted and what that specific user is already permitted to do. The smaller set always wins. Key behavioral characteristics: - Default-deny posture: nothing is accessible until explicitly granted - Grants are structured hierarchically: org-level, then team-level, then per-member overrides - Off-limits tools are not surfaced to the agent at all — they are excluded from the tool list built per user - Permissions are read live on each request, so policy changes take effect immediately without redeployment - The same agent deployment behaves differently for different users without requiring separate agent instances Permissions can be scoped down to individual actions within a tool, not just entire tools. Reusable "scope sets" allow administrators to define common permission bundles (e.g., "support tools, read-only") and apply them across multiple contexts. The system is intended for organizations deploying shared AI agents across teams with varying access levels, ensuring that an agent can never exceed the permissions of the person it is acting for.