Amazon SageMaker AI introduces AI agent experience for streamlined model customization

Amazon SageMaker AI has unveiled a new agentic experience designed to significantly expedite the process of model customization. Previously a task that could take several months, this new workflow reduces the time required to just days or even hours. This advancement is particularly beneficial for customers developing AI solutions, as they must meticulously define their use case goals and success criteria, prepare data, select appropriate models, and conduct multiple experiments with different models and fine-tuning techniques. Once an optimal model is identified, determining the most cost-effective deployment strategy becomes crucial.

The newly introduced capability allows developers to interact with coding agents using natural language, streamlining the entire process from defining the use case to deploying a high-quality model in production. This agentic experience, leveraging SageMaker AI model customization skills, offers expertise in fine-tuning tailored to a developer’s specific use case. It also includes data transformation into required formats, comprehensive quality evaluation using LLM-as-a-judge metrics, and flexible deployment options to Amazon Bedrock or SageMaker AI endpoints.

Customers have the flexibility to install these skills in any Integrated Development Environment (IDE) of their choice, including Visual Studio and Cursor. Developers can collaborate with various coding agents such as Kiro, Claude Code, and CoPilot to optimize popular model families like Amazon Nova, Llama, Qwen, and GPT-OSS. The experience also generates reusable and editable code artifacts, ensuring transparency, reproducibility, and the potential for automation through integration with AIOps pipelines.

To utilize these capabilities, users can install the SageMaker AI skills in their preferred IDE via the sagemaker-ai agent plugin. The skills are also pre-installed in SageMaker Studio Notebooks, alongside the Kiro coding agent. Users simply need to subscribe to Kiro, open the chat window in Studio Notebooks, and engage with the agent to construct their workflow.

The experience supports advanced customization techniques, including supervised fine-tuning for instruction tuning, direct Preference Optimization for adjusting tone and preference selections, and Reinforcement Learning for use cases requiring verifiable correctness. For additional information on model customization with the AI agent experience in Amazon SageMaker AI, interested parties are encouraged to consult the SageMaker model customization documentation.