CivBench ran agents 300+ turns through Civilization VI and found they stop checking state before losing.

CivBench exposes 76 MCP tools and a narration layer converting game state to structured text, with a single episode spanning 300+ turns and thousands of tool calls. Across 23 admissible runs the authors decline to rank models and instead introduce Proactive Monitoring Rate and RAG@10 (whether commitments stated in planning reflections get executed within ten turns). Despite playbook instructions to query victory progress every 20 turns, agents queried every 30 to 75, and in 7 of 20 detectable defeats never queried inside the warning window. Agents under-monitor state that's available but must be explicitly requested.