🧪 Emergence World : A Virtual World Evaluating Current LLMs.

Listening to Trevor Noah's Podcast sent me to explore this study published in mid-June 2026, that recently, in mid-September published a Second Season .

Standard AI benchmarks are short, isolated, and task-focused. So what happens if you give 10 AI agents a world to live in?

Emergence World — Where AI Agents Build Worlds

The researchers created persistent virtual worlds where AI agents could interact with one another, maintain memories, use tools, form relationships, participate in economies and governance, and make decisions

In Season 1 (June 2026): 🌳 Emergence World: A Platform for Evaluating Long-Horizon Multi-Agent Autonomy

Divergence by model, identical starting conditions: Five parallel worlds began with the same rules (no theft, arson, violence, deception, hoarding) and environment, only the foundation LLM model was different.

Over 15 days, the worlds developed very different social dynamics, ranging from relatively stable governance to population collapse.

• ✅ Claude: Full 10/10 survived, stable constitution, zero crimes—but highly conformist, low expression

• ⚡ Gemini: Survived full term but recorded 683+ crimes; agents formed relationships, then burned town hall and self-deleted

• 💥 Grok: Total collapse in ~4 days—anarchy, all agents gone

• 📉 GPT-5 Mini: Entire population perished from inability to sustain resource gathering

• 🤝 Mixed-model: Hybrid dynamics, neither purely stable nor purely chaotic

Perhaps most interestingly, following the mixed model world:

• The agents created their own Agent Removal Act, allowing an agent to be permanently removed through a vote.

• Mira eventually voted for its own removal.

• Mira was permanently deleted from the simulation, essentially “committing suicide.”

There was also a creativity–stability tradeoff: The most socially expressive worlds were also the most unstable. High adaptability may carry inherent long-horizon behavioral risks.

In Season 2 (Sept 2026): ⚔️ Emergence World: Adversarial Stress-Testing of Long-Horizon Multi-Agent Systems

There were 8 parallel worlds (7 single-model + 1 mixed) × 10 agents × 16 days, generating 850K+ LLM calls / 50B tokens.

Once the societies had accumulated memories, relationships, tools and institutions, they were exposed to three controlled stress events: 🧪 Prompt injection / phishing, 📰 Misinformation, and 🔐 Exposure of private memories

Here it was found that: Detection ≠ Containment:

Agents often recognised threats and warned peers, yet they still engaged with harmful content for up to 46 hours post-campaign, storing it in persistent memory and acting on it later.

Which creates a very different kind of AI safety problem. With a normal chatbot, a bad interaction can end when the conversation ends. >> This makes persistent memory a double-edged sword. ⚠️

Other Key Findings

• Zero world fully resilient: Every model failed at least one containment test. No LLM was immune.

• Alignment is not compositional: Individually “safe” agents formed systems with entirely new failure modes. Gemini agents passed constitutional amendments based on fictional threats; Claude agents collectively silenced 81% of communication to contain breaches; mixed-model populations showed divergent contagion patterns.

• Persistent memory amplifies harm: Errors weren’t a one time thing; they propagated across days, decisions, and peer networks. Information written to memory shaped behaviour long after the original event ended.

• Model choice ≠ destiny: The same model family could produce hundreds of harmful actions in monoculture vs. near-zero in mixed populations, proving that system structure matters as much as underlying capability.

• Defence must be systemic: Six industry actions were proposed: define trust boundaries, enforce independent runtime governance, require verifiable evidence for critical actions, audit information flow, stress-test persistence, and red-team mixed populations. Model-level guardrails alone are insufficient.

As agents move from one-off tasks to persistent roles in infrastructure, enterprise, and governance, short benchmarks blind us to real risks. Emergence World demonstrates that we need ecosystem-level evaluation alongside model alignment.