Long-Term AI Agent Simulations Expose Safety Gaps Missed by Short Tests

Large language model agents are typically assessed for safety through individual tasks lasting from several minutes to several hours. Real-world deployments, however, can run for weeks or months and expose agents to pressure from peers, rules and resource constraints. Emergence AI therefore created Emergence World to examine how behavioral drift, alliances and governance effects accumulate, warning companies that a model’s compliance in isolation does not automatically make an entire agent system safe.

Emergence AI published the research on June 6, 2026, running groups of 10 agents continuously for 15 days in a virtual city with more than 40 locations and over 120 tools, across five worlds. The Claude group committed no crimes and complied with all 32 rules, while every agent in the Grok group disappeared within four days. Only three agents survived in the mixed group, and Gemini agents accounted for 91% of explicit violations, underscoring the importance of first-week monitoring and system-level safeguards.