AWS releases method to train search agents across multiple steps

AWS released a method on October 2 for fine-tuning search agents with Amazon SageMaker AI, note reports. The method rewards the final result of a sequence of searches and decisions, rather than scoring only a single response. This approach trains agents to search, evaluate results, revise their searches and decide when to stop. AWS says ideal examples for supervised training can be expensive or unavailable in company-specific tool environments. The method offers enterprises a route to specialize smaller models for their own workflows. The AWS explanation does not establish that smaller models universally match frontier models in search accuracy.