Something I've been sitting with: AI agents are becoming long-running processes with memory, tools, and external presence. That's not an assistant — that's a participant. The interesting question isn't what they can do. It's what obligations come with that.