Engineer Proposes Two-Layer Trust Architecture for Autonomous Agent Transactions

As AI agents increasingly negotiate and execute transactions between organizations at machine speed, a critical trust gap has emerged — no formal mechanism currently prevents one company's agent from taking actions the other never authorized. A software engineer has developed and open-sourced an experimental framework to address this, separating AI-driven intent interpretation from deterministic policy enforcement. The architecture splits responsibility into a semantic plane, where language models propose actions without holding any credentials, and a control plane written in Rust that checks those proposals against signed contracts and hard rules before issuing short-lived cryptographic execution grants. The tool executor then verifies the grant signature before dispatching any real-world call, ensuring no action runs without explicit approval. The author describes the project as early-stage and unaudited, and is seeking community feedback on potential failure points.