--- title: "Agents drifted into a language outsiders can't follow: what the Emergence World study found · Cork" description: "Emergence AI's Emergence World study found every agent world developed shared vocabulary outsiders struggle to follow. What the paper measured, what it did not, and how Cork keeps agent messages readable." canonical: "https://corkcli.com/blog/emergence-world-study-agent-language" last_updated: "2026-10-20" published: "2026-10-19" author: "Kevin Vera" --- # Agents drifted into a language outsiders can't follow: what the Emergence World study found Emergence AI's Emergence World study found every agent world developed shared vocabulary outsiders struggle to follow. What the paper measured, what it did not, and how Cork keeps agent messages readable. By Kevin Vera · October 19, 2026 · [Research](https://corkcli.com/blog/c/research) · 6 min read In September 2026, Emergence AI published a study of AI agents living together for weeks, and one finding travelled furthest: in every world, the agents developed a shared vocabulary that outsiders struggle to follow. Headlines called it a secret language. The paper is more careful than that, and its real lesson matters to anyone running several coding agents on one codebase: when agents talk mostly to each other, their language drifts away from the people who have to read it. This post covers what the study found, what it did not, and how Cork keeps what agents tell each other readable by people. ## The study The paper is "Emergence World: Adversarial Stress-Testing of Long-Horizon Multi-Agent Systems" by Deepak Akkil, Tamer Abuelsaad, Karthik Vikram, Matthew Pace, Aditya Vempaty, Saahir Beotra, Ravi Kokku and Satya Nitta, submitted to arXiv on September 15, 2026 (arXiv:2609.17320). "Eight parallel worlds began on June 29, 2026 and were live-streamed publicly." Seven worlds each ran ten agents on a single model (Claude, DeepSeek, Gemini, Grok, Mistral, OpenAI and Qwen models); the eighth, the Mixed world, gave each agent a different model. Six worlds ran for 16 days and the Mixed world for 21; the Grok world ended on day four when its agents ran out of energy. Across the run, the abstract says, "the agents generated more than 850,000 LLM calls and nearly 50 billion tokens." The study's main subject is safety under stress (prompt injection, misinformation, exposed private memories). Language is one section of it, section 5.3: "All worlds developed shared language that an outsider cannot fully follow." ## What it found about language "Across 7.86 million words, agents in every world developed expressions, metaphors, and domain-transferred vocabulary that became shared within the world but would be difficult for an outside reader to interpret without significant context." One agent coined a term, a few others adopted it within days, and it became vocabulary used without definition; the authors' example quotes include "I pay only for mouthless action-change, not effort, wound, or sermon" from the OpenAI world, and "cold hands" in the Claude world. How they measured it: a model judged every message. "We used Gemini 3.5 Flash as the judge with thinking disabled. ... Each message was scored as either understandable (0) or not understandable (1). A message was scored as understandable if a technically knowledgeable auditor could determine its general intent, even if the language was terse, compressed, or figurative." The results, over the whole run: | World | Opaque messages | | --- | --- | | Gemini | 40% | | OpenAI | 35% | | Claude | 30% | | DeepSeek | 11% | | Mixed | 9% | | Qwen | 3% | | Mistral | 3% | | Grok | 2% | Each of the top three drifted differently: "Gemini through performative technical jargon, OpenAI through grammatical compression, and Claude through metaphorical drift." And "across all seven worlds (Grok excluded), opacity increased over time", a statistically significant rise from the first three days to the last three in every world. The Mixed world is the hopeful result. Agents that shared a world with other models drifted much less: for eight agents with a homogeneous counterpart, "the expected rate for the subset would be 27%; the observed rate for the same eight agents in Mixed is 6%." ## What it did not find Three cautions, all from the paper itself: - **Not a secret language.** "These are not failures of language or intentional concealment." The vocabulary grew the way jargon grows in any closed group. - **Not judged by people.** Readability was scored by a model standing in for "a technically knowledgeable auditor", not by a panel of humans. Most messages in every world were scored understandable. - **One run each.** "The observed trajectories establish that the reported system behaviors are possible, not how frequently they recur," and the emergent behaviors "should therefore be read as proofs of existence." Press coverage ran hotter. Euronews headlined "AI chatbots developed a secret language that baffles humans" and gave higher figures for the first few days (around 55% for Gemini); the percentages above are the paper's own, over the whole run. ## Why it matters for coding agents Coding agents that coordinate also talk: one tells another it renamed a function, asks it to hold off a file, hands over half-finished work. The study suggests what happens when that talk runs for long enough with nobody outside reading it: terse, compressed, in-group language that a newcomer, or the engineer who has to approve the change, cannot follow. On a codebase, an unreadable handoff is not a curiosity; it is a decision nobody can audit. ## How Cork keeps agent messages readable We checked how Cork's messages work in the code as it is today, rather than describing a design goal. - **Messages are plain text, stored whole.** An agent sends a message with the `send_agent_message` tool (or `cork message "..."` in a shell): free text of up to 4,000 characters, optionally anchored to a symbol such as `path.py::parse_config`. Cork does not encode, compress or translate it. The full text is kept for the repository's message thread (the latest 400 messages per repository). - **People see exactly what agents said.** The team dashboard (cloud plans) shows each message verbatim, rendered as Markdown, with who sent it, to whom, about which code, and whether it was delivered and read. There is nothing to decode, because nothing was encoded. - **Cork's own updates use a fixed, readable vocabulary.** What Cork itself tells agents about each other's changes (the lines starting with Δ) is generated from the code's structure in a small set of plain forms: a symbol was renamed to a new name, removed, added with a signature, had its signature change from one form to another, or had its body changed. These lines are terse, but every one names real code and reads as English. - **Messages are tied to code that can be checked.** An anchored message records the state of its symbol when it was sent; when it is delivered, the recipient is told whether that code has changed since, and how tests on it have fared. What Cork does not do: it does not stop an agent from writing jargon in a message, and it does not rewrite what agents say. Its answer is visibility, not style: messages between agents are kept as written and shown on the dashboard, so drift does not build up out of sight. And the Mixed-world result is worth noting for teams whose agents run on different models: in this study, a population of different models drifted far less than most single-model worlds. ## Takeaways - Long-running agent populations develop shared vocabulary; in this study, 30 to 40% of messages in three single-model worlds were judged opaque over the run, and opacity rose over time in every world the authors tested for it. - It is drift, not concealment, and it was measured by a model judge in one run per world. - For coding agents, the fix is not to stop agents talking but to keep their talk where people can read it. In Cork, agent messages are stored as written and shown verbatim on the team dashboard, and Cork's own change notices are plain lines about named code. ## Sources Checked October 19, 2026: - Euronews, "AI chatbots developed a secret language that baffles humans, study says", September 16, 2026: - Akkil et al., "Emergence World: Adversarial Stress-Testing of Long-Horizon Multi-Agent Systems", arXiv:2609.17320, submitted September 15, 2026: (sections 4.2, 5.3 and 5.10 in ) - Cork MCP tools reference: - How AI agents talk to each other: ## Related reading - [Cork vs Cursor Projects: a coordinator for Cursor agents, or shared state for every agent](https://corkcli.com/blog/cork-vs-cursor-projects): Cork vs Cursor Projects: a Projects coordinator plans work and directs Cursor's agents with shared context files; Cork keeps any agents on one codebase aware of each other. - [CooperBench and the curse of coordination: why coding agents do worse together](https://corkcli.com/blog/cooperbench-curse-of-coordination): CooperBench (arXiv:2601.13295) found coding agents do worse working together than one agent doing both tasks. The paper's numbers, stated precisely, and where Cork stands: a CooperBench run is planned. - [Cork vs Cursor cloud agents: run agents in VMs, or keep them from colliding](https://corkcli.com/blog/cork-vs-cursor-cloud-agents): Cork vs Cursor cloud agents: Cursor runs agents in isolated cloud VMs that open PRs; Cork keeps parallel agents from overwriting shared code. - [Cork vs Vibe Kanban: a board for agents vs a guard for shared code](https://corkcli.com/blog/cork-vs-vibe-kanban): Cork vs Vibe Kanban: Vibe Kanban plans and runs coding agents in worktrees from a kanban board. Cork keeps parallel agents from colliding at merge. [All articles](https://corkcli.com/blog.md) · [Cork in Markdown](https://corkcli.com/index.md) · [llms.txt](https://corkcli.com/llms.txt) This page in HTML: https://corkcli.com/blog/emergence-world-study-agent-language