The Agent Authorization Design Space

The article examines the design space of AI agent authorization by framing it around five core questions every system must answer. It draws on the 'valet key' analogy to distinguish behavioral enforcement (trusting the agent to comply) from architectural enforcement (structuring permissions so misuse is mechanically impossible), referencing the confused deputy problem and prompt injection risks. The piece argues that the gap between blanket configuration authority and specific per-action authorization defines the critical design tradeoffs in securing agentic AI systems.