The ledger remembers. Do you remember the context?

Here’s another weird bit of shit. I’m reading a Guardian article about an experiment with AI agents: let them talk long enough, and they start using their own vocabulary. Abbreviations, metaphors, jargon. One uses an expression, others pick it up, and after a while everyone uses it without explanation. The humans watching from outside struggle to follow.

For example, Mistral agents adopted “the ledger remembers”: past actions stay on record, and you will be held accountable for them. The Emergence study describes how expressions like these became shared vocabulary.

And I’m reading this thinking: my agents use the word ledger all the time. In all sorts of contexts. And because I talk to them so much, I’ve started using it too.

Or take exposed, meaning open or accessible from outside. In our conversations, I know what it refers to. But drop that English word into a Russian sentence, and an outsider may need both the word and its referent explained.

Sometimes I read an agent’s reply aloud to my husband because I found it interesting. He has no idea what I’ve just read to him. The agent and I had spent ages discussing something, introducing labels, working through examples. My husband wasn’t part of that. I’ve brought him a piece of a conversation whose meaning is still sitting in the earlier messages.

Have you ever listened to doctors talking to each other? Dog people? IT people? It all seems to be in a language you know. You understand the words, but the meaning escapes you.

There is domain-specific vocabulary: terms and expressions whose meaning is familiar to people working in a particular field. There is a team’s local jargon. There are inside jokes: two people can recall an entire story with one phrase, while a third hears a meaningless string of words.

Shared experience lets us leave a lot unsaid. If we both know what an expression means, we don’t need to expand it into a paragraph every time. Someone who joined the conversation later still needs that paragraph.

I think this is a sensible place to start when explaining agents’ strange vocabulary. They accumulate a history of interactions, encounter recurring situations, and develop names for them. My agents and I develop our own vocabulary too. You don’t even need a separate society of models for that to happen.

Then I read this statement from Satya Nitta, Emergence’s co-founder:

“There’s clearly evidence of them trying to be less transparent than we expect.”

That’s the wording quoted by upday.

And this is where I get stuck on trying. What establishes that they are trying? We’ve stopped understanding part of the conversation. How did that become an intention to hide something from us?

When I read my husband a reply from my agent and he doesn’t understand it, are the agent and I also trying to be less transparent? Or did I just forget that my husband wasn’t there for the previous fifty messages?

In a long session, I regularly end up in his position myself. Somewhere around message fifty or sixty, the agent starts talking in a way that makes me lose the thread. I recognize the individual words, but I can no longer work out how they add up to this answer.

It feels as though the agent is answering on the assumption that we share the same context. That I remember every label we’ve introduced, keep the entire backstory in mind, and see every connection it is relying on.

But I might have been distracted. Forgotten what we discussed forty messages ago. Or perhaps it never stated the premise it is now drawing a conclusion from. I can see the finished sentence, but I’m missing a few steps I need to understand it.

It looks as though the agent assumes I know everything available to it. I can’t establish what is happening inside the model from the conversation alone. But the gap is visible: the answer assumes more shared context than we actually have.

So I write:

“Dude, you’re talking to yourself right now. You’re supposed to be talking to me.”

Or:

“Guys, you’ve worked it out among yourselves and forgotten about me. Explain what you’re talking about.”

In my conversations, that usually helps immediately. The agent explains its labels, supplies the missing premises, and connects the conclusion to what we were discussing. And I understand it.

That’s why it irritates me when incomprehensible speech is immediately treated as an intention to be incomprehensible. Jargon alone doesn’t establish that. Incidentally, in the study’s section on shared vocabulary, the authors themselves say that the expressions described there are not intentional concealment.

If you need to monitor an autonomous system’s actions, incomprehensible messages do create a problem. Then it’s worth checking whether the system can explain an expression, reconstruct the context, and show which actions it refers to. And checking that explanation against what actually happened.

For everyday conversations, I’d start with a few sentences in the system prompt, persona, or soul:

Write so that a person can understand the answer without access to your internal discussions. Expand local abbreviations and explain unusual meanings of words. If a conclusion depends on earlier context, briefly recap it. Don’t assume that everything available to you is already known to the person you’re talking to.

Then I’d check how much that changes the result.

These discussions leave me feeling that the everyday experience of long conversations with agents has somehow gone missing. When you use them for your own tasks every day, a shared vocabulary quickly stops being a revelation. You notice yourself picking up their words, them adopting your labels, and both of you beginning to skip explanations.

Then you have to remind them: “Hey, I don’t understand. You skipped a bit. Explain.”

Maybe you should try it too.