MetaRoute-Bench: A New Framework for Evaluating Agentic Workflow Routing Policies

MetaRoute-Bench has been introduced to evaluate how decisions are made in agentic systems, which often encounter choices like whether to answer questions directly, break tasks down, use tools, run code, consult experts, verify results, or recover from mistakes. These decisions impact how well tasks are completed, their costs, and the time taken, but they’re typically only assessed through overall task accuracy. This benchmark provides a clear and open way to analyze these decision-making processes within a single execution model. It includes 180 synthetic task profiles in areas like data analysis and document processing, and features eight routing policies and 30 random seed pairs. A task-aware policy achieved a 79.4% success rate, surpassing other approaches. The benchmark is available on arXiv with the identifier 2608.00107.