Amazon adds agentic fine-tuning assistant to SageMaker AI

On May 5, 2026 Amazon updated SageMaker AI with an agentic fine-tuning assistant that takes plain — language use cases, prepares data, runs training, and returns editable Jupyter notebooks for Llama, Qwen, Deepseek and Nova.

On May 5, 2026 Amazon introduced an agentic workflow assistant inside SageMaker AI designed to simplify customization and fine-tuning of large language models. The assistant accepts plain — language descriptions of a desired use case and orchestrates the end-to-end fine-tuning pipeline within the SageMaker development environment, reducing the need for engineers to assemble data pipelines and training steps manually.

The agent recommends appropriate training approaches, prepares and validates datasets, launches training jobs and delivers finished code as Jupyter notebooks. All generated code is editable and reusable, so developers can modify the notebooks, iterate locally and reproduce or extend the pipeline without starting from scratch. By producing working notebooks rather than opaque artifacts, the assistant aims to make the output immediately actionable for engineering teams.

Amazon’s Kiro AI agent is preinstalled in the SageMaker environment, but customers can replace it with alternatives such as Claude Code or with their own custom agents. The platform includes nine prebuilt “skills” that cover stages from dataset validation to model deployment, effectively codifying common fine-tuning tasks so teams do not have to reimplement routine workflow steps each time they experiment.

The new feature explicitly supports multiple model families — including Llama, Qwen, Deepseek and Amazon’s Nova-enabling teams to run agent — guided fine-tuning across third — party and Amazon models. That cross — family support is intended to remove the friction of hand-translating APIs or formats between toolchains, so projects can move from specification to training without time-consuming compatibility work.

For builders, the combination of plain — language specification and an agent — managed pipeline is intended to lower the overhead of trying different models and training strategies. By handling orchestration and returning editable notebooks, SageMaker AI seeks to accelerate iteration cycles: teams can inspect generated code, tweak preprocessing or fine-tuning parameters, and relaunch experiments more quickly than when wiring pipelines by hand. Taken together, the additions aim to make production — ready fine-tuning workflows more accessible to engineering teams. The agentic assistant codifies recurring workflow steps, standardizes validation and deployment tasks through its prebuilt skills, and provides tangible, editable outputs that developers can adapt for their specific use cases.