Where do you answer"is the agent allowed to do this?"–one place,orevery adapter?
The GitHub repo "warden" by YogiSotho introduces a unified authorization system that centralizes the question "is the agent allowed to do this?" into a single place, rather than scattering the answer across every adapter or component.
Background
- This is a GitHub issue/readme about a software project called **Warden**, an authorization layer for AI agents (LLM-based autonomous programs). The core question: should you check "is this agent allowed to do this?" in one central place, or in every adapter (i.e., every tool/integration the agent calls)?
- **Why it matters**: AI agents today can browse the web, send emails, execute code, or control APIs. Without a single authorization point, permissions are scattered across dozens of tool adapters, making it easy to miss a security check or create inconsistent rules. A "PDP" (Policy Decision Point) centralizes this, so one policy engine decides every action.
- **Prior context**: This mirrors the "policy as code" movement in cloud/DevOps (e.g., OPA, Cedar). The author is proposing Warden as a PDP for agent systems, letting developers write a single policy (e.g., "agent can read files but only in /tmp") rather than hard-coding checks into each adapter.