AgentGrad: Intervention-guided Prompt Optimization for Multi Agent Systems

Summary

The paper identifies two weaknesses in textual-gradient prompt optimization for LLM multi-agent systems: target prompts are chosen without verifying that modifying them resolves the failure, and gradients are derived without agent-level supervision; aggregation then concatenates randomly grouped gradients, mixing unrelated failure modes. AgentGrad tackles both. Sequential intervention injects a ground-truth-derived hint into one agent at a time, in reverse execution order; the agent whose correction restores maximum reward becomes the target, and its hint-adjusted intermediate output serves as an agent-level pseudo-label for extracting a fine-grained textual gradient. Semantic textual gradient abstraction then clusters sample-level gradients into 'semantic minibatches' sharing a corrective pattern (with a cyclic soft cluster-size bound, e.g., 5→3→1) and abstracts each cluster into one generalized gradient; candidate prompts are validated first on their minibatch, then on held-out validation data. Across five benchmarks (HotpotQA, HoVer, PUPA, IFBench, MATH), the authors report the best averages with GPT-5-mini (+11.76 points over the unoptimized baseline) and Qwen3-8B (+9.67), plus 2.5× lower mean wall-clock time than the next-fastest baseline, GEPA (136 vs 337 minutes), and the best transfer to five unseen same-domain benchmarks. Ablations credit target identification, agent-level supervision, and abstraction separately. Key limitations: three seeds with standard errors only and no significance tests; the 'state-of-the-art on both backbones' claim conflicts with Table 2, where TextGrad beats AgentGrad on IFBench/Qwen3-8B (42.52 vs 41.42); hints require ground truth; no code release is stated.