Fine-tune a search agent with multi-turn RL on Amazon SageMaker AI

This article describes how to fine-tune LLM-powered search agents using Amazon SageMaker AI multi-turn reinforcement learning (MTRL) to optimize multi-turn interactions and improve retrieval quality.

• MTRL trains agents across full multi-turn trajectories rather than single steps, optimizing interdependent decisions in search workflows
• Modular agent-environment interface enables custom rewards, tool loops, and conversation shapes with low-code integration
• Serverless execution with per-token pricing eliminates GPU cluster provisioning and management overhead
• Fine-tuned Qwen3.6-27B model achieved +23.7% nDCG@10 gain on BrowseComp-Plus and reduced failure rate from 22.89% to 0.68%
• Minimal configuration required: only three hyperparameters changed from defaults (max_epochs, global_batch_size, rollout_max_concurrency)
• Trajectory and reward observability via MLflow enables inspection of agent behavior turn-by-turn across training steps

Amazon SageMaker AI MTRL provides a practical, low-barrier approach to training specialized search agents with improved reliability and efficiency without deep RL expertise.