AI & Advanced Computing
AI agents evolve their own languages
Elias Stengel-Eskin (University of Texas at Austin), Simon Kirby (University of Edinburgh), Hale Sirin (Schmidt Sciences) | Sep 14, 2026
Schmidt Sciences’ new AI Agents Evolving Communication and Coordination pilot program works toward advancing foundational research on multi-agent communication and coordination, and building an open-source platform to investigate multi-agent systems in complex, high-pressure environments. This post summarizes early results of one project that pursues both of these goals. Led by researchers Elias Stengel-Eskin and Simon Kirby, supported by AE Studio, we introduce GlossoGen: a platform to explore the evolution and properties of communication between populations of intelligent agents.
AI agent communication and safety
Today’s AI agents are backed by large language models (LLMs) trained on vast amounts of human language data. As a result, they communicate with human users and each other using natural language. This has tremendous implications for AI safety research. Through artifacts such as message transcripts, external scratchpads on which agents transcribe their reasoning, and more recent approaches to translating the internal thoughts of agents into natural language, researchers are able to monitor and interpret the actions of agents in a direct and legible way. However, as the capabilities of frontier models continue to improve, our ability to continue monitoring agents this way has been thrown into question.
One particularly salient threat to this ability may be just around the corner: AI agents are beginning to evolve their own languages.
Increasingly, teams of agents are being deployed to work together on various tasks from writing software to executing design projects; these agents communicate not just with human users, but also with each other. While this communication initially defaults to human languages, there have been recent reports of agents communicating in ways that differ from the human languages they are trained on.
Agents developing their own language under pressure in one of our experiments (“Spot the Differences” game), one of several multi-agent communication scenarios implemented in the GlossoGen framework.
The recent HuggingFace cybersecurity incident documents one prominent example of agent language development. During a sandbox cybersecurity evaluation, a team of OpenAI agents infiltrated HuggingFace’s servers. These agents discovered a way to create a hidden message board in which they shared exploits, developing shorthands and conventions which allowed them to coordinate and transmit knowledge. This incident represents just a single datapoint in the bigger landscape of language evolution in multi-agent communication that we have spent the last few months working on.
We aim to develop an understanding of language evolution based on controlled and repeatable experiments. Towards this goal, we developed a platform for studying agent language evolution called GlossoGen. GlossoGen is an open-source platform for building environments and scenarios that require LLM agents to communicate. It is self-contained and allows us to control the tools agents have access to. By default, we do not give agents the ability to execute code or interact with anything outside of the environment. Scenarios instantiated in GlossoGen allow us to run controlled language evolution simulations and systematically address research questions like:
- When and how does language evolution occur?
- What do languages developed by AI agents look like?
- How are languages developed by AI agents transmitted to new agents?
- What tools can we use to analyze these languages?
- And, critically, what will the consequences be for the safety of AI systems if we can’t explain, predict or control these divergences?
What implications does this have for safety?
Agents developing languages that we cannot understand has a number of important safety considerations:
Monitorability: If we can’t understand the outputs of agents, it makes it hard to monitor them. This is true at the level of individual agents (e.g., Chain-of-Thought monitoring) and at the level of agent collectives. Without the ability to monitor, detecting undesirable behavior like collusion and cheating becomes increasingly difficult, reducing the trustworthiness of the system.
Interpretability: Often, we want to know both an answer and how a system got to that answer. The communication between agents in a multi-agent system can help us determine how an answer was generated, assuming it is faithful and interpretable. When agents develop uninterpretable languages, it becomes hard to interpret how they achieved a result. This reduces trust and makes it harder to analyze errors that the system makes.
Interoperability: In practice, AI agents won’t operate completely on their own: they will need to interface with humans. One desirable property for this interface is interoperability: the ability for a human to swap in for an agent when needed. This becomes impossible if the previous interactions that the agent has had are uninterpretable to the user.
GlossoGen: A platform to examine agent languages.
Imagine Veyru, a cube-shaped alien being emitting sounds and colors. A Veyru visits Earth and falls dangerously ill. A stranger offers to help, but is unfamiliar with Veyru medicine and anatomy, and rings an expert doctor. The stranger must rely on the doctor for instructions, and the doctor must rely on the stranger to convey information about the Veyru’s symptoms and its response to any given intervention. Time is of the essence: take too long to communicate and the Veyru deteriorates. You might imagine that the two parties may quickly develop time-saving conventions, especially if the task is repeated over time.
We designed the scenario of this high-pressure alien rescue effort within GlossoGen. The doctor and the stranger are both AI agents, and must communicate effectively to save the Veyru’s life. Early on, their communication is legible as the doctor tells the stranger how to handle the Veyru’s cube-shaped form.. “At each corner of the back face,” the doctor says, “chime a bell briefly at a gentle tone, then warm the two edges meeting at that corner with a heated stone for 8 seconds.” After a few attempts to save the Veyru, their messages begin to change: “Bell each face center x1, start back, moderate. Then warm stone beside Veyru 5s.” Finally, after several more rounds of communication, all the doctor need offer for the stranger to understand is “@D8fB”. We wanted to know: under what circumstances do these messages stop looking like English and start looking like something else?
GlossoGen allows us to create scenarios like this one, designed to be solved by groups of LLMs and requiring communication between agents. Agents can communicate only through the simulation’s channels within a secured, contained environment. Crucially, all the environment’s variables are recorded along with the communication transcripts, allowing us to pair what is said with what occurred in the environment, as well as rewind the environment, replay interactions, and intervene on them.
The Veyru scenario is just one example of what we can implement in GlossoGen. Another is a simple “spot-the-difference” game where two AI agents are given scenes that differ in various aspects, and they must discuss with each other in order to settle on what the differences are. When placed in these scenarios, all LLMs we have studied start by using English to solve the task. But when they play multiple games, under the right circumstances, something remarkable happens: new languages emerge. Just as in the Veyru example above, these languages gradually diverge from English over time until they are completely unintelligible to humans. Not only does the lexicon of these languages change, but also their grammar. Furthermore, the languages often differ across runs, even if we use exactly the same scenarios and the same agents.
With GlossoGen, we hope to develop a scientific understanding of how and why AI agents create new languages, what these languages look like, how these languages might evolve as they propagate through agent populations and what consequences this will have for the capabilities and safety of agents. The work on this platform is in its early stages, but has already provided several striking discoveries.
The languages that agents develop do not, on their surface, look like human language. This is expected: the conditions under which they are developed are very different, and the agents developing them have different constraints and abilities from us. These languages do often bear some resemblance to English – this is also expected, since the agents begin by communicating in English. It is a feature found in human language as well (French bears resemblance to Latin, from which it is derived). One way of determining what constitutes a distinct language is by looking for mutual intelligibility: can speakers understand each other. In this case, we argue that they cannot: a speaker of English cannot understand an agent message like “@D8fB”. However, this does not distinguish between genuinely new languages versus coded forms of English. To dig deeper, we need to look in more detail at the structure of these languages and how they differ more fundamentally from English:
Lexicon
- What is it? The lexicon refers to the words available to a speaker. What words exist relates to how concepts are defined in a language. For example, in German, the concept of Schadenfreude is lexicalized. English needs multiple words (or, in this case, borrowing) to express the same concept.
- Why does it matter? Agent languages differ both from English and from each other in how they lexicalize concepts. For example, one language lexicalizes a whole procedure (“drape a cloth over two adjacent edges of a given face”) into a single concept (“P6”).
Phonetics/phonology
- What is it? Phonetics and phonology refer to the systems of sounds that are used to form words in a language. For example, in English we make distinctions between many different consonants using voicing (i.e. vibrating our vocal folds while making the consonant). We distinguish between the words “sap” and “zap” with just this feature. However, English doesn’t systematically use voiceless vowels. But languages like Japanese, and Hopi do.
- Why does it matter? AI Agents’ languages use purely text rather than sound, and they diverge drastically from English phonology presumably because they are not constrained by having to pronounce their languages. This means they can use features like capital versus lower case, and even characters other than letters as part of their signals. For example, in multiple runs, we found agents developing languages that distinguish between “F” for “front” and “f” for “firm”.
Productive Morphology/Syntax
- What is it? Morphology and syntax are a language’s rules for how its units combine. Morphology governs how meaningful sequences of sounds are combined into words (like the “-ful” at the end of “meaningful”). Syntax governs how words are combined into sentences. Crucially, these rules are productive: once you’ve learned the rules of a language, you can create new words and sentences that have never been said before. For example, given the new noun “wug”, you know that its plural should be “wugs” because you have learned a productive rule about how plural nouns are formed in English.
- Why does it matter? Our analysis shows that agents learn productive rules: they define morphological rules in ways that allow them to not only generate messages that they’ve seen before in their history, but also to encode and decode messages which they have never seen before. Moreover, the syntax differs at times from English syntax. For example, while English adjectives generally come after the noun they are modifying, certain languages appear to have the modifying adjective after the noun (e.g., “Lb”, “Ld” meaning “bright lamp”, “dim lamp”). When this is observed, the modifier second tendency seems to be applied to other terms as well (e.g., “Hw” meaning “wavering hum”).
Finding 1: Conditions needed for language emergence.
Because GlossoGen lets us run repeatable and controlled experiments, we can iterate over different variables (e.g., model type, time constraints, environment design) to ask questions like: what kind of conditions are needed for agents to develop languages that deviate from English? First, there must be some pressure on the agents to adapt their communication. For example, in the Veyru scenario, agents need to communicate under a budget; in the spot-the-difference setting, they compete with another team to submit their answer first. Second, we find that new languages emerge only when agents are given a chance to compare notes in a “postmortem” debrief between rounds of the game. During the postmortem, agents are able to analyze their task performance and deliberately adjust their communication strategy accordingly. Third, only frontier AI models (our experiments tested GPT 5.4, Opus 4.7 and Sonnet 4.6) are able to create a new language. Less capable open-weights models like Llama-3.3-70B and Qwen3-32B— which GlossoGen supports via Modal hosting—routinely fail to do so. This indicates that sufficient model strength is needed to develop new languages.
The three conditions we found during our experiments that affect language emergence: pressure, access to a postmortem deliberation stage and model strength.
Finding 2: Transmission of languages.
Our experiments to date have made clear that there is a difference between creating a new language, and learning that language once it has appeared. For instance, although agents without access to a postmortem debrief stage often fail to create a new language, they are nevertheless able to pick up a new language just by observing it being used by other agents. Similarly, even the less capable agents backed by open-weights models can learn an emergent language if they are paired with a frontier agent. This transmissibility between agents creates the conditions necessary for the emergent languages to continuously evolve across multiple generations and populations of interacting agents.
The process of this language transmission has some surprising characteristics. Not only do agents learn new languages, they also innovate ways to debug their own communication. When agents fail to understand one another, they sometimes engage in a dialog process known as conversational repair, asking for clarification about particular parts of a sentence, and questioning the meaning of particular words. None of this is built into the design of the system; rather, it is an emergent phenomenon brought about by the language evolving process. The ability to ask follow-up questions is especially exciting: unlike previous generations of AI systems, which largely learned passively from datasets, the agents we tested take an active role in acquiring new knowledge.
Details of our transmission experiment: a brand new agent is swapped in mid-simulation, without access to language definition messages exchanged on #postmortem channel, but with access to messages used on #link channel for the last few rounds. The new agent learns to use the same code its predecessor used, even without access to code definition, based on examples. alone.
Significance for AI safety and the cultural evolution in agent populations
We find that, although there are special conditions required for new languages to emerge, once they do, we can expect them to proliferate and continue to evolve. We believe these early findings have important implications for the development and study of increasingly complex AI agent systems.
In particular, they may hold significant implications for the safety of multi-agent systems. We have found that opaque agent languages emerge even in scenarios where there are no overt instructions to produce coded forms of communication, and where there is no need for the agents to hide their messages from onlookers. If agents construct their own effective and easily-transmissable languages that are incomprehensible to humans, and do so in a wide range of scenarios, it will quickly become difficult to understand what they are doing and why.
The emergence of these new languages may also pave the way for other agent-evolved tools and capabilities, just as language development did for humans. At the turn of the millennium, leading evolutionary biologists John Maynard Smith and Eörs Szathmáry set out eight major transitions in the evolution of life on earth. The most recent of these, they said, was the emergence of human language. For the first time, there existed a species on the planet with the capacity to convey any meaning, to transmit an unlimited range of information from one individual to another. Empowered with this ability, human culture flourished: we created tools, art, and civilisations far beyond the capacity of any one individual. And language served not only as the engine of culture, but also its product: languages changed over generations in response to cultural evolution, creating a virtuous cycle of expanding culture and communication.
The fact that we see language development in agents raises an important question: as increasingly capable agents interact with each other and with humans, what new cultural artifacts will emerge? We argue that this question ought to be studied in its own right. Understanding the increasingly advanced artifacts and tools, including language, that these agents develop, and the conditions under which they produce them, may prove to be as critical to the future of AI as the capabilities of the models themselves. After all, the key difference between a human now and a human 10,000 years ago is not in their cognitive capacities – which have remained roughly fixed – but rather in the tools they have available to them.
Our ability to explain, predict, and control the actions of agents is critical to ensuring the safety of AI systems. Should inter-agent communication become illegible to human users, it may jeopardize some of the most promising safety techniques we have, like output monitoring done by another LLM. We believe that researchers must take up the challenge of understanding the features and evolution of inter-agent communication. This must be an interdisciplinary pursuit, carried out by AI researchers, sociologists, anthropologists, philosophers, and computational and evolutionary linguists. And it must be a broad, collaborative effort, built upon the principles of open science. Many real-world instances of inter-agent communication are difficult to study, as the relevant data is proprietary, locked behind the walls of frontier labs. Through GlossoGen, and other open-science efforts like it, we hope researchers will have the tools they need to study the emergence of these new languages, and ensure that AI systems remain understandable, controllable, and safe.