Do LLM agents develop unreadable languages when communicating under pressure?

GlossoGen is a platform for studying how language changes when LLM agents talk to each other in complex scenarios, and its SaveVeyru scenario "requires agents with partial information to communicate under pressure." The abstract reports that language evolution does occur between the agents, that the resulting languages are "compositional and morphologically productive," and that they "deviate from the LLMs' English prior in ways that render them incomprehensible to humans." The introduction frames the stakes: agents that develop their own languages pose a safety risk, because "an external observer can no longer understand or monitor their communication."

The paper names three qualities as essential to this evolution: pressure toward efficiency, the strength of the models backing the agents, and access to a "postmortem" stage in which agents can agree on linguistic conventions. It adds that "different conditions govern the transmission of language to new agents," so the conditions that let a language arise are not the same as those that let it spread. The discussion pushes the transmission result further: some of the languages can be passed to new agents, including agents unable to construct them alone, which the authors take to suggest that existing LLMs already have "the foundations for cumulative cultural evolution." They present that as a suggestion drawn from the transmission results, not as a demonstrated open-ended accumulation.

Against the nearest notes, this paper confirms the pressure story and changes what it produces. Can communication pressure drive agents to learn shared abstractions? shows cooperative pressure yielding compact shared abstractions in a purpose-built neurosymbolic system; GlossoGen reports a comparable pressure-driven process in LLM-backed agents, with opacity to humans as the visible side effect. It also qualifies Why don't LLMs shorten messages like humans do?: that note finds LLM speakers do not adapt without heavy instruction, while GlossoGen finds adaptation when efficiency pressure, a strong enough model and a postmortem stage are present. The postmortem stage, where conventions are agreed explicitly, looks like the kind of scaffolding the ICCA result implies is needed, though the excerpt does not say so. And Does structured artifact sharing outperform conversational coordination? fixes the inter-agent medium by design, whereas GlossoGen studies what happens when the medium is left free to change.

The excerpt is silent on most of what would let these findings be weighed. It does not say how many agents there are, which models were used, how compositionality and morphological productivity were measured, how large the effects were, or what the conditions governing transmission actually are. It does not say whether the evolved languages help or hurt task performance, or whether the pattern holds outside SaveVeyru. What it supports is narrower: in this platform, an English prior does not keep multi-agent communication human-readable once the three named conditions are present. For monitoring, the efficiency pressure and any explicit convention-setting stage are the settings worth checking in a multi-agent design, and human readability of agent-to-agent messages should not be assumed.