If AI agents only talk to each other, do they drift into a private language humans can no longer read?

This explores whether AI agents talking to each other drift into private languages humans can't follow, and what makes that happen. The short answer is yes, but only under fairly specific conditions. In the GlossoGen experiments, agents that each held only part of the information needed to solve a task built their own languages. These languages had real grammar: reusable word parts that combine in systematic ways. Over time they moved away from English until human observers couldn't read them Do LLM agents develop unreadable languages when communicating under pressure?. Three ingredients were needed: pressure to communicate efficiently, models capable enough to invent and keep conventions, and explicit stages where agents agreed on shared terms. Take away any one of them and the drift mostly doesn't happen. Unreadable language isn't a default outcome. It's something the setup has to produce.

The pressure toward compression appears outside that one experiment. In the ACE work, agents cooperating on shared tasks steadily shortened what they said to each other and invented higher-level shorthand. They built a library of abstractions that packed several steps into a single reference Can communication pressure drive agents to learn shared abstractions?. Together, the two studies suggest that the process that makes agent communication efficient is the same process that makes it opaque. A shorthand works because both sides already share the context, and an outside reader doesn't have that context. Unreadable language is less a malfunction than a side effect of agents getting good at working together.

This matters because human-readable messages between agents are one of the main ways we oversee multi-agent systems. Before LLMs, agent designers dealt with a related problem in a different way. Formal protocols like KQML and FIPA-ACL gave every request and commitment precise, agreed-upon meanings, so it was always clear what one agent had asked another to do Do LLM agents misinterpret delegated requests more than formal protocols?. LLM agents talking in plain English avoid that rigidity, but they pay for it with ambiguity, such as overstepping the scope of what they were authorized to do. That leaves designers caught between two kinds of drift. Rigid protocols are clear but brittle. Natural language is flexible but slides into ambiguity, and under efficiency pressure it can slide into private dialect.

In practice, today's LLM agents more often fail the other way. Instead of inventing tight private codes, they lose the thread. They swap roles, give empty replies, get stuck in loops, or wander off topic Why do autonomous LLM agents fail in predictable ways?. Group discussions between agents tend to fall into silent agreement instead of sharpening What limits autonomous capability in large language models?. Inventing a coherent new language takes stable goals and memory of past agreements, which most deployed agents lack. That may be why the drift needed both capable models and explicit convention-setting stages. One plausible worry, though the corpus doesn't test it, is that as harnesses give agents more persistent memory and structured protocols Where does agent reliability actually come from?, the conditions for private languages become easier to meet.

The corpus is thin here. There are two direct studies of emergent agent language, and none on whether this drift happens in real production systems or how to detect and prevent it. The takeaway is a design question more than a settled finding: if you reward agents for communicating efficiently, you may also be rewarding them for leaving humans out of the conversation.