AgentBaton Evaluates Model Handoffs for Long-Horizon Agentic Tasks

This study investigates model collaboration through handoffs between capable cloud models and efficient local models to reduce task costs. Findings indicate that efficient-to-capable model handoffs generally underperform compared to using capable-only models for maintaining task accuracy.