A control room for agent societies where the live intervention command and the replay evidence are the same object
GOD (arXiv 2608.27992, Aug 28) addresses the gap between a finished replay and raw logs when a run contains many agents, messages, commands and model calls. Its technical claim is a command-and-artifact loop where live controls and replay evidence share one operator command model, so asking a question and intervening in a run use the same interface, with package contracts separating scenario, map and profile data from local runtime state so a run travels to another researcher intact. Evaluated on 15 completed run slots: across 14 intervention runs, 78 of 84 target-agent checks recorded the commanded destination and 169 of 182 state answers matched a saved location or action string. The sample is small and the domain is generative-agent simulation, so treat the architecture as the takeaway rather than the numbers.
↳ Follow the thread