Developer Proposes Two-Layer Architecture to Limit AI Agent API Risks

A developer experimenting with agent safety has identified a risky pattern in modern AI agent setups, where language models are handed API tokens and can directly execute backend calls without independent oversight. The researcher proposes separating reasoning and execution into two distinct layers: a probabilistic 'Semantic Plane' where AI models interpret intent but hold no credentials, and a deterministic 'Control Plane' built in Rust that validates proposed actions against hard policies before granting execution rights. The Control Plane issues short-lived, single-use cryptographic grants, ensuring no action reaches a backend service without passing structured checks. The approach aims to prevent issues such as prompt injection, hallucination-driven misfires, and runaway recursive calls between collaborating agents. The author stresses this remains an early-stage personal exploration and has not been formally audited or proposed as an industry standard.