At a Glance
Under time pressure and with a post-task review, modern language models acting as teammates often invent compact, non-English codes that improve task performance but become unreadable to humans, creating monitoring and trust risks.
ON THIS PAGE
What They Found
Agents backed by strong language models begin to replace plain English with shorter, task-specific codes when communication is costly and they can deliberate afterward. These emergent codes are structured (Reflection Pattern), they follow reusable grammar-like patterns, help teams succeed under tight time budgets, and can be learned by new agents that only observe usage (Semantic Capability Matching Pattern). Weaker open models generally do not invent such codes on their own but can pick them up through observation, meaning codes can spread even if not every agent can create them.
Not sure where to start?Get personalized recommendations
Data Highlights
1≈430% increase in mean perplexity under an English model (from 320 to ~1700) when agents had tight budgets and postmortem access versus high-budget/no-postmortem settings.
2120 total runs (3 proprietary models × 2 budgets × postmortem on/off), with 10 instances per setting and 15 rounds per run.
3High-pressure budget = 150 seconds versus low-pressure = 2000 seconds (time cost implemented as 1 second per character), which strongly correlates with language change.
What This Means
Engineers building multi-agent systems and platform teams should care because emergent compressed codes can boost throughput but break human monitoring and debugging. Safety, compliance, and product leads should audit communication channels and post-task logs to detect unreadable conventions and manage trust risks, guided in part by the Human-in-the-Loop Pattern.
Key Figures
Fig 1: Figure 1: Example of language emergence under budget pressure in an emergency response scenario (example from Sonnet 4.6). In our fictional SaveVeyru scenario, a Field Observer has access to an ailing alien, Veyru. The field observer reports observed symptoms to the Specialist, who can provide remedies which the Observer implements. Agents must handle information asymmetry, forcing them to communicate, and time pressure, encouraging them to develop novel languages, as seen in Round 13.
Fig 2: Figure 2: Sketch of Veyru scenario and its information asymmetry. Information about the Veyru is only observed directly by the Field Observer, who can take actions on the Veyru. Information about the Stellar Oracle is observed only by the Specialist.
Fig 8: Figure 8: We show the paradigm saturation with (right) and without (left) semantic-role pooling. Saturation measures how many of the potential codes have been observed in the link channel. High saturation means out of all of the possible codes our paradigm would permit, a large percentage of them are attested. Without pooling, we see high levels of saturation, suggesting that almost all of the codes which are permitted by the paradigms have already been generated, thereby limiting our ability to generate and test novel forms of codes. With pooling, however, we gain the ability to test a wide variety of novel codes.
Ready to evaluate your AI agents?
Learn how ReputAgent helps teams build trustworthy AI through systematic evaluation.
Learn MoreYes, But...
The experiments use a fictional emergency-response task designed to prevent models from relying on stored medical knowledge, so real-world effects may differ. Results are strongest for the proprietary models tested; weaker open models didn't create new codes though they could learn them, so model strength matters. The study focuses on communication pressure and post-task review as drivers—other environmental factors might change outcomes and were not exhaustively explored. Governance and evaluation considerations are relevant, as highlighted by the Evaluation-Driven Development (EDDOps).
Methodology & More
GlossoGen is a configurable environment for running multi-agent interactions between language models, built to study how inter-agent communication changes over time. The team implemented a fictional emergency-response scenario called SaveVeyru: one agent (Field Observer) sees detailed symptoms and acts; another (Specialist) knows how to remedy but cannot act directly. Agents exchange messages through a Slack-style channel where each character costs time; experiments vary the time budget (150s vs 2000s) and whether agents get a postmortem deliberation stage after rounds. Runs used a mix of proprietary and open-weight models, with 10 instances per configuration and 15 rounds per run. Under tight time budgets and when a postmortem deliberation channel is available, stronger models consistently develop compressed, non-English-like codes that reduce message length and increase task success. These emergent codes show productive structure—a kind of morphosyntax—that lets agents generate and interpret novel forms beyond their training examples. In swap experiments where a newcomer only observes usage (not the postmortem), agents can learn these codes from context, showing transmissibility. The findings imply that multi-agent systems can exhibit cumulative cultural evolution: innovations arise, stick, and spread—helpful for efficiency but risky for monitoring, traceability, and trust unless teams add detection and governance mechanisms. Planning Pattern
Avoid common pitfallsLearn what failures to watch for
Credibility Assessment:
Lead author has h-index ~19 (recognized researcher level). However it is an arXiv preprint with mixed low-profile coauthors and no venue peer review.