Bilevel Game-Theoretic Framework for Multi-Agent LLM Coordination

This framework models orchestrator-worker interactions as a bilevel coordination game to provide a unified account of task decomposition and reflection. It treats reflection as stochastic movement over semantic memory and provides finite-time upper bounds for free-form reflection performance.