Amazon SageMaker AI opens custom training sessions

The launch builds on SageMaker AI's existing serverless model customization, which already offered managed techniques such as multi-turn RL. AWS says customers asked for more control — to define their own loss functions, shape reward signals, or chain RL stages in one loop — and training sessions are the response.

The design targets a specific pain point in agent training: a loop that samples responses on one set of infrastructure, scores them on another, and feeds results back can require connecting separate tools and clusters. With training sessions, the customer's code runs the training loop and reward function, while SageMaker AI performs the GPU-heavy work of sampling and gradient updates.

Early users describe demanding use cases. ImagineSoftware used the feature for a healthcare claim-denial triage model, citing direct control over forward and backward passes and the optimizer step. Karini AI fine-tuned gpt-oss-20b for multi-turn tool calling against its Amazon Neptune knowledge graph, with its CEO calling checkpoint-and-resume reliable. AWS also says customer data and customized weights stay in the customer's S3 buckets and are not used to train another model — a detail likely to matter to regulated industries.