Agent-guided workflows to accelerate model customization in Amazon SageMaker AI

Amazon SageMaker AI has introduced a new agent-guided workflow that transforms how organizations customize foundation models with proprietary data. Instead of navigating fragmented APIs, mastering fine-tuning techniques like Supervised Fine-Tuning (SFT), Direct Preference Optimization (DPO), and Reinforcement Learning from Verifiable Rewards (RLVR), or managing months-long experiment cycles, developers now describe their use case in natural language. The AI coding agent activates purpose-built Skills—pre-built, modular instruction sets covering data preparation, technique selection, hyperparameter configuration, model evaluation, and deployment. These Skills encode deep AWS and data science expertise, generate ready-to-run notebooks at each step, and produce fully editable code that integrates into existing pipelines. The agent also optimizes token usage, lowering costs while maintaining quality.

This experience is natively integrated into SageMaker AI Studio JupyterLab, where the default Amazon Kiro agent or any ACP-compatible agent (e.g., Claude Code) automatically loads relevant Skills from the workspace. Developers can customize each Skill to match their team’s governance standards and tooling preferences, enabling reproducible best practices across the organization. The generated code can deploy models to either Amazon Bedrock or SageMaker AI endpoints, providing flexibility. With nine initial Skills covering the full customization lifecycle, the feature eliminates the need for deep ML expertise, making fine-tuning accessible to a broader range of teams while accelerating experimentation and production deployment.