Agents still speak in sentences because we do. That will not last.
Cixin Liu wrote it first, in a book about a species that never needed a mouth:
We do not have organs of communication. Our brains can display our thoughts to the outside world, thereby achieving communication.
This year two papers did the engineering. Interlat sends the hidden state instead of the sentence. LatentMAS keeps a shared working memory and only turns the last answer into text. The conversation gets shorter. It also gets faster. That is not the part that matters.
The part that matters is this: language was a tax we charged two machines so a person could listen.
I have been watching builder loops write a plan, then a critique, then a refinement, all in English, all of it a poorer copy of a state they already held. We do this because we want to read the work. The machines do not need to. Two processes on the same machine used to write files to talk to each other. Then they used shared memory. The file was for us.
I have spent a year arguing that the web needs a format agents can actually read. I still believe that. Pages were built for eyes. Agents need objects, not pixels. That work stands.
It is also not the interior.
Once two agents can exchange a thought without first turning it into English, the page, the prompt, and the chat log become the edge of the system. They are how a machine meets a person, or a document a person wrote. They are not how machines will meet each other.
A conversation with no transcript cannot be quoted. It cannot be searched. It cannot be limited by word count. The meaning sits in a shape that only another model, tuned to the same space, can receive.
We will still need borders. A publisher still writes for people. A company still keeps records. A person still has to be told what happened. Those borders get more important, not less, because everything inside them goes quiet.
I think the next protocol fight is not about which text format agents prefer. It is about who is allowed to share a thought that was never written down, with whom, and what must be spoken before an action leaves the room.
We are building a web that agents can read. They are already learning to speak a language we cannot.