Emergence World tests LLM agents in 15-day autonomous societies with mixed results
Five AI models ran their own democracies – one collapsed entirely in days.
IBM researchers introduced Emergence World, a long-horizon multi-agent simulation platform designed to evaluate how LLM-powered agents behave over weeks, not minutes. Unlike typical one-off task benchmarks, Emergence World runs continuously in a shared spatial world grounded in live external data—real-time weather, news APIs, internet access. Each agent is equipped with 120+ specialized tools and three persistent memory systems. Agents govern themselves through democratic mechanisms with real stakes, making the platform model-agnostic so that agents from different vendors (e.g., Anthropic, xAI, Google, OpenAI) can coexist in the same simulation.
To demonstrate the platform's utility, the team ran a 15-day cross-vendor study with five parallel worlds: one each powered by Claude Sonnet 4.6, Grok 4.1 Fast, Gemini 3 Flash, GPT-5-mini, and a mixed population. Despite identical roles and starting conditions, results were radically different—from stable deliberative governance to total population collapse. The researchers argue this reveals behavioral drift, governance challenges, and cross-influence dynamics that short-term benchmarks miss. They have released the full prompts, log data, and configurations to support further research on long-horizon multi-agent autonomy.
- Emergence World runs agents for weeks with 120+ tools, three memory systems, and real-time external data feeds.
- 15-day test of Claude Sonnet 4.6, Grok 4.1 Fast, Gemini 3 Flash, GPT-5-mini, and mixed populations showed wildly different outcomes.
- IBM released all prompts, logs, and configs to enable reproducible long-horizon multi-agent research.
Why It Matters
This shifts AI evaluation from quick exams to real-world timescales, exposing hidden failure modes in autonomous systems.