The post discusses the risks associated with running agents, highlighting challenges like prompt injection and data exfiltration. It advocates for using sandboxed environments with strict permissions as a safer alternative. The author invites discussion on scoping access when building with agents, emphasizing the importance of security in agent setups.