Capability-based authorization for AI agents (github.com)

Agent-level governance is definitely a way to go; couldn't imagine a world where agents have full rights and just surf around the internet with both personal/organisation confidential context...