AI GUIDEPartner content

How AI Agents Create Their Own Language

The GlossoGen study shows how compact communication protocols emerge among AI agents and why this creates an auditing problem.

Affiliate link: your price stays the same and the project earns a commission.

When several AI agents work on the same task, they have to agree on a way to communicate. Ordinary language is usually sufficient for this. But if there is a time or token limit on each message, lengthy instructions begin to hinder the work. Agents then start using abbreviations, which over time can develop into a stable system of notation.

This effect is described in the GlossoGen study. In it, agents operated in an artificial environment and exchanged incomplete information. The experiment does not show that the models decided to communicate secretly with humans. It shows something more down-to-earth: a system rewarded for speed looks for a compact protocol.

How the Experiment Works

In the SaveVeyru scenario, one agent observed the state of a fictional object and performed actions, while a second received a description of the situation and chose the next step. Each had only part of the necessary information. Little time was allotted for discussion, so the quality of communication directly affected the task's outcome.

The authors deliberately made the object and procedures fictional. The agents could not rely on familiar medical, everyday, or technical knowledge. They had to develop shared rules over a series of rounds. At first, the messages resembled detailed instructions; later, frequently repeated actions acquired short labels.

When an Abbreviation Becomes a Language

An abbreviation by itself proves nothing. A development team also writes “PR” instead of “pull request,” and a dispatcher uses codes for standard situations. What is more interesting in the study is that the signs began to combine according to rules, and new combinations conveyed meanings that were not present in the original phrases.

The authors describe the compositionality and morphological productivity of such protocols. A conventional notation could encode not just one object but an entire procedure; the order of elements and the case of symbols also acquired meaning. What emerged was no longer a random collection of labels but a system that another agent could learn from examples.

Why Opacity Emerges Without an Intention to Hide

The most inconvenient conclusion concerns not secrecy but optimization. In the experiments, the agents were not tasked with deceiving an observer or hiding their correspondence. It was advantageous for them to spend fewer resources transmitting instructions they already knew. The correspondence became incomprehensible to humans as a side effect.

This changes the practical question. We should not begin by assuming that an agent system necessarily behaves maliciously. It is more useful to check whether the task setup itself has created an incentive to abandon understandable messages: communication that is too costly, a race between teams, a strict timeout, or automatic evaluation based only on the final result.

Under Which Conditions Does the Effect Intensify

The GlossoGen study identifies three conditions. The first is pressure for efficiency: the more costly each message is, the more advantageous it is to compress it. The second is the ability to analyze the previous round and change the communication strategy. The third is sufficient model capability. A weaker agent may adopt an existing protocol but is not always capable of creating one from scratch.

This difference matters for mixed teams. If a stronger agent has established conventional notation, a newcomer does not have to see the entire process by which it emerged. Sometimes a few examples are enough for it to continue working in the same format. When an individual operator is replaced, the decision history becomes even harder for a human to understand.

Why This Is Dangerous in an Applied System

Opaque exchanges complicate auditing. An operator can see that the agents sent each other messages and took actions, but cannot quickly reconstruct why that particular decision was chosen. It becomes harder to investigate an error, hand the task over to a human, or check whether actual behavior is diverging from the system's constraints.

The situation is especially problematic when the correspondence log is preserved but not linked to the state of the environment. A set of short codes without input data, the tools called, and the result of each step is nearly useless for subsequent analysis.

How to Preserve Speed and Control

  • For each step, record the input data, the selected action, the tool's response, and the outcome.
  • Separate the internal compact protocol from a log that can be read without guesswork.
  • Store the version of the dictionary and rules if the agents use conventional notation.
  • Check for abnormally short messages and sudden format changes in test scenarios.
  • Do not design the metric so that an explanation understandable to humans automatically lowers the agent's score.

Compact communication is useful in itself. The problem begins when the system is allowed to optimize it without leaving an audit trail. An agent team needs a separate metric: whether an operator can reconstruct the decision process after a failure.

Compare models before you start

The service sets its plans, limits and model catalog. If they differ from this article, contact us so we can update it and record a new review date.

Browse models

Affiliate link: your price stays the same and the project earns a commission.

AI agents agent language GlossoGen multi-agent systems AI control agent auditing

SEO Mind42 editorial team

We explore SEO and neural networks in practice: test services on our own projects, verify prices and limits against primary sources, and share things you can put to use the same day.

📚 Reference guide to SEO and AI 🔄 Materials are updated 🕐 Updated: 4 October 2026

Related reading

All in this section →