AWS launches agent-guided workflows in SageMaker AI to automate model fine-tuning

Amazon Web Services has released agent-guided workflows in SageMaker AI that use AI coding agents to automate the model customization process from data preparation through deployment.

Nine pre-built skills

The system includes nine modular skills built on the Agent Skills open format:

1. Use Case Specification - Structured discovery for business problem definition
2. Planning Discovery - Generates multi-step customization plans
3. Fine-tuning Setup - Selects base models from SageMaker AI Hub and recommends techniques
4. Dataset Evaluation - Validates dataset format and schema
5. Dataset Transformation - Converts between ML data formats (OpenAI chat, SageMaker AI, Hugging Face, Amazon Nova)
6. Fine-tuning Training - Generates training notebooks for serverless fine-tuning
7. Model Evaluation - Configures LLM-as-Judge evaluation with built-in and custom metrics
8. Model Deployment - Determines deployment pathway and generates code
9. SageMaker API Integration - Calls SageMaker AI APIs, accesses S3 data sources, and interacts with model registries

Supported fine-tuning techniques

The workflows support three fine-tuning methods:

• SFT (Supervised Fine-Tuning): Trains on input/output pairs for task-specific behavior, instruction following, and domain adaptation
• DPO (Direct Preference Optimization): Trains on preferred versus rejected outputs for aligning tone, style, and subjective preferences
• RLVR (Reinforcement Learning with Verifiable Rewards): Uses code-based reward functions for tasks where correctness can be programmatically verified

The system recommends the appropriate technique during the planning phase based on the use case.

Agent integration

SageMaker AI JupyterLab now includes integrated support through the Agent Communication Protocol (ACP). Amazon's Kiro agent comes pre-configured in the chat panel by default. Users can also configure other ACP-compatible agents including Claude Code, Cursor, and similar tools.

When coding agents operate within SageMaker AI JupyterLab, the environment automatically loads relevant model customization skills into the agent's context. All generated code is fully editable and produces reusable artifacts.

Requirements

To use the feature, organizations need:

• An AWS account with SageMaker AI domain access
• An AWS IAM role with required permissions
• An Amazon S3 bucket
• SageMaker AI Studio JupyterLab compute space
• SageMaker AI Distribution image version 4.1 or higher
• AmazonSageMakerFullAccess managed policy attached to the domain's execution role
• Additional inline policy for Lambda, S3, and Bedrock access
• Trust policy allowing sagemaker.amazonaws.com, lambda.amazonaws.com, and bedrock.amazonaws.com to assume the role

The feature has no minimum instance type requirement.

What this means

AWS is productizing the agent workflow pattern that emerged with tools like Claude Code and Cursor, but with domain-specific expertise baked in. The nine pre-built skills address a genuine pain point: teams that understand their use case but lack deep knowledge of SageMaker APIs, fine-tuning techniques, or AWS service integration patterns. By making these skills customizable, AWS enables organizations to encode their own governance standards and workflows rather than relying solely on general-purpose coding assistants. The approach demonstrates how cloud providers are moving beyond raw infrastructure to offer opinionated, automated workflows for common ML operations.