AgentGrad: Intervention-guided Prompt Optimization for Multi Agent Systems
AgentGrad tests prompt fixes one agent at a time before turning failures into reusable gradients.
The paper says existing textual-gradient methods can target the wrong agent and merge unrelated failures. AgentGrad uses sequential intervention to find which agent change actually resolves a failure, then uses that changed output as supervision. It also clusters similar gradients before abstracting them into broader corrective patterns. The authors report state-of-the-art results on five multi-agent benchmarks and a 2.5x average reduction in optimization time versus the next-fastest baseline. HF Daily Papers' note
The paper says existing textual-gradient methods can target the wrong agent and merge unrelated failures. AgentGrad uses sequential intervention to find which agent change actually resolves a failure, then uses that changed output as supervision. It also clusters similar gradients before abstracting them into broader corrective patterns. The authors report state-of-the-art results on five multi-agent benchmarks and a 2.5x average reduction in optimization time versus the next-fastest baseline. HF Daily Papers' note
score 5