AgentGrad is an intervention-guided prompt optimization method for LLM-based multi-agent systems. It addresses limitations in existing textual-gradient approaches during gradient extraction and aggregation, where natural-language feedback guides prompt updates.