I wrote an essay about what agents need at their foundation in order to enable to the possibility of securely designed tool delegation. It’s easy to be insecure. What would it take to be secure?

https://ryanhiebert.com/oauth-for-ai-harnesses/