New Benchmark for Agentic Workflow Routing Improves Efficiency

Researchers have developed a new executable benchmark and a sophisticated meta-router to address the complex decision-making required in agentic AI systems. These systems often need to choose between various operations—such as direct answering, task decomposition, evidence retrieval, code execution, or delegation—to effectively complete a task. Current routing methods typically focus on isolated choices, but this new approach allows for the composition of diverse operations. The benchmark includes a wide range of tasks across data analysis, research, and document processing, with outcomes machine-checked for accuracy. The budget-aware meta-router learns to predict operation probabilities and greedily composes them under cost and action-count constraints. On held-out test data, the learned policy achieved a 100% success rate, outperforming strong static workflows (93.5%) and reducing costs by 43%. While performance dropped on a lexically shifted challenge split, it still demonstrated significant cost savings and outperformed one-shot routing. The findings highlight lexical generalization as a key limitation, establishing a reproducible testbed and a proof of concept for advanced agentic routing.