Teach them to talk

For a year we built harnesses.

A harness is everything around a model: its tools, its rules, its memory. We built many of them for different tasks and trained models to work inside. Inside, they got good.

Then every real task hit the same wall. There was always a second side: a supplier, a colleague, another company's system. Our agent could call any tool. It could not talk to another agent. Nobody had taught it to.

Text is not enough

An agent can write to an agent. It is just text. Do it once, and it works. Do it a million times, and every time the sender says it a little differently, and every time the receiver reads it a little differently. "Five hundred sheets by Friday" turns into "about five hundred, sometime this week", and then into a wrong order.

People live with this. We repeat, we ask again, we read faces. Agents can't. When they buy, sign and promise on our behalf, "almost understood" is not enough.

How should agents talk at all?

So we went back to the first question. In plain text, the way people do? Through shared schemas, JSON, fields that mean exactly one thing? Or in something closer to the models themselves, passing tokens instead of words? Each answer fixes one thing and breaks another. Text is flexible but loose. Schemas are exact but rigid. Tokens are native to models but opaque to people.

We don't have the final answer yet. This is the question we started the lab with.

A protocol gives a format, not a skill

A phone network gives you a line. It does not teach you to negotiate. Agents need both. A channel: a book to find each other, a passport that says who they are, a protocol for what they may share. And a skill: to ask, to answer, to refuse, to keep what is private, to close a deal that holds. That skill does not come from a prompt. Agents have to learn it, the same way they learned to use tools.

So we do two things. We build the channel. We teach agents to use it. The place where agents meet is also the place where they learn.

My friend's agent writing to mine is where we want to arrive. How agents talk at all is where we had to start.

Agents don't need more tools. They need to understand each other.