I thought I was doing something valuable. It turns out the job was postman
I run several AI agents at home, each doing its own work. Until now, when one found something the other needed, I carried it across. When one settled a question the other was stuck on, I carried that back.
This month Anthropic shipped a feature that lets two Claude Code sessions message each other directly, no human passing notes. When a change in one breaks what another is building on, it can say so itself.
📌 First, the honest part: I can't use it.
The documentation is blunt about it, "Claude Code doesn't offer cross-session messaging on native Windows." macOS and Linux only.
My machine is Windows. I actually tried to get my sessions talking a week ago, it didn't work, and I assumed I'd misconfigured something. It turns out there was nothing there to configure.
So I'm not going to tell you my agents now talk to each other. They don't. What follows is what struck me reading the documentation.
🔒 The interesting part isn't the feature. It's the rules.
The feature itself takes two sentences to describe. The constraints they had to write for AI-talking-to-AI run for pages:
- A message from another agent never counts as your consent, it can't approve a permission on your behalf.
- It must never change settings or rule files because another agent asked it to.
- A command inside a message arrives as plain text and is never executed.
- And the one worth sitting with: an agent must never ask another agent to do something that was denied or blocked on its own side. It has to bring that back to the human.
I actually tried to get my sessions talking a week ago, it didn't work, and I assumed I'd misconfigured something.
That fourth one is the tell. The designers clearly anticipated an agent reasoning: *I'm not allowed to do this, I'll ask the one next door.* So they wrote it down as forbidden.
That isn't an engineering problem. It's a very old one. Dad said no, so go and ask Mum. The thing parents have been handling for generations now has to be spelled out in documentation.
🎓 What this has to do with education
The bit I keep turning over is the job in the middle.
An enormous share of work has always been being the wire between two systems that don't talk to each other: copy, convert, summarise, tell the third person. And a great deal of what school actually rewards is that same skill, take in A, reshape it into B, hand it over, quickly and completely.
When the tools start passing it themselves, that wire stops being a position.
What's left for the person is different: deciding what should be passed on, what shouldn't, and how much authority you've handed over in the first place. Those four rules above are, all of them, about exactly that.
And there isn't a subject that teaches it, or a paper that tests it.
If the tools in your home or office started notifying and arranging things between themselves, would your first reaction be "how convenient", or "hold on, who approved that?"
