CivBench tests 300-turn Civilization VI agents

CivBench runs agents through 300-plus turns, thousands of tool calls, and 76 MCP tools; across 23 pilot runs, agents executed only 48.2% to 65.8% of commitments they made in their own planning reflections, according to the CivBench paper. Agents also queried strategic state every 30 to 75 turns despite guidance to check every 20, missing that warning window before 7 of 20 detectable defeats. Builders should treat agent infrastructure benchmarks that price each task as incomplete without state-monitoring and commitment-follow-through tests before granting long-horizon autonomy.