AWS SageMaker AI Gets Agentic Fine-Tuning Workflows with 9 Built-In Skills and Kiro and Claude Code Integration

AWS presented agent-guided workflows in Amazon SageMaker AI on May 4, 2026 — a system offering a conversational approach to model customization through nine built-in agentic skills, with direct integration of Kiro and Claude Code in the JupyterLab environment. The goal is to reduce the deep ML expertise requirement for fine-tuning work to a natural-language problem description, with the agentic system generating editable Jupyter notebooks at each step of the pipeline.

How does an agent-guided workflow get started?

A developer describes their use case in natural language through the chat panel in SageMaker AI Studio JupyterLab (e.g., “I need a clinical reasoning model that interprets medical cases step by step”). The coding agent identifies relevant skills, activates them sequentially, and at each stage generates an editable Jupyter notebook that the developer can modify before execution. AWS claims in the announcement that agent skills “not only increase productivity but also reduce token consumption” through targeted activation.

The nine skills cover the entire customization lifecycle: Use Case Specification, Planning/Discovery, Fine-tuning Setup, Dataset Evaluation, Dataset Transformation, Fine-tuning, Model Evaluation and Model Deployment, plus an orchestration step. The system supports SFT (Supervised Fine-Tuning), DPO (Direct Preference Optimization) and RLVR (Reinforcement Learning with Verifiable Rewards). Supported model families: Amazon Nova, GPT-OSS, Llama, Qwen and DeepSeek.

What is the difference between the Kiro and Claude Code integrations?

Kiro is the default agent — preconfigured in the SageMaker AI Studio JupyterLab chat panel and authenticated via device flow. Claude Code is installed via the npm package @zed-industries/claude-agent-acp, supports ACP (Agent Communication Protocol) and integrates with Amazon Bedrock through a configuration file. Both agents automatically access SageMaker skill agents in the JupyterLab environment.

The technical architecture relies on ACP compatibility, integration with the SageMaker AI Hub (base models), Amazon S3 (data storage), MLflow (metrics tracking) and Amazon Bedrock (deployment). The demo use case in the announcement is a clinical reasoning model that “walks through medical cases step by step before reaching a diagnosis” — an example that directly correlates with the ReClaim foundation model trend for medical applications (see the parallel arXiv article from the same day).

What does this mean for enterprise ML teams?

AWS’s claim that “what once required months of specialized ML work can now be completed in days” is significant — but it remains verifiable only once teams produce their first production models through the workflow. Bigger picture: AWS is positioning SageMaker as an integrated agent-orchestration platform, similar to IBM’s simultaneously announced next-gen watsonx Orchestrate and AWS Bedrock AgentCore Optimization (launched the same morning). The convergence toward “agent-guided model customization” as the standard enterprise interface is now a clear industry trend, not an experimental approach.

Pricing was not announced in the launch post, suggesting standard SageMaker billing usage-based pricing. Availability is for all organizations with AWS accounts and a SageMaker AI domain — no geographic restrictions noted in the announcement.