Can AI agents invent their own private language to work together, and what has to be true for that to happen?
What specific conditions allow language evolution in multi-agent LLM systems?
This explores what has to be true for a group of LLM agents to invent their own working language instead of chatting in English, and what the library knows about it.
This explores what has to be true for a group of LLM agents to invent their own working language instead of chatting in English. The library has one direct study, and it names three conditions. In the GlossoGen platform, each agent holds only part of the information needed to solve a task, so they have to talk. Under those terms they developed languages with productive word-building that drifted away from English and became opaque to human observers. That took efficiency pressure (talk had to be economical, so shorthand paid off), models capable enough to pull it off, and explicit convention-setting stages where agents settled how they would talk Do LLM agents develop unreadable languages when communicating under pressure?.
The convention-setting stage matters most. The language didn't emerge from casual back-and-forth. It needed a deliberate step where agents agreed on terms, and agreement is the thing other notes in the collection show LLM groups are bad at. Agent groups often fail to reach consensus because they stall or time out, not because someone corrupts a value, and success drops as the group grows even with no bad actors Can LLM agent groups reliably reach consensus together?. At larger scale, agents also either settle on a strategy too late or adopt one without telling their neighbors Why do multi-agent systems fail to coordinate at scale?. A shared language is a consensus problem, so it should get harder as the group grows. That is my inference. Nothing in the collection tests it for invented languages.
The same coordination note shows the other side of the coin. Agents accept neighbors' information without checking it. That habit would spread a new convention quickly, and it would spread a mistaken one just as quickly.
The capability condition is a threshold, and the notes suggest it isn't a simple more-is-better scale. LLM agents in cooperative settings flip roles, send empty replies, loop endlessly, or wander off topic, because they lack a persistent goal and a stable role identity Why do autonomous LLM agents fail in predictable ways?. A convention only works if agents keep hold of it and of who they are in the conversation, so models that can't do that fail before language can form. A neighboring finding on agent harnesses (the scaffolding around a model) shows the same shape: models gain the most from agreed-upon structure at mid-tier capability, and the strongest models struggle to follow instructions faithfully Do stronger models always evolve harnesses better?. That study is about harnesses, not language, so it hints at the pattern without confirming it.
The evidence is thin. Everything specific here comes from one platform, so the three conditions describe what worked in one setup, not a proven recipe. The collection also doesn't say whether these languages stay stable, transfer to new tasks, or can be decoded by humans afterward.