What's New

Amazon SageMaker AI has launched an agentic experience for model customization that dramatically accelerates the end-to-end fine-tuning workflow, compressing what previously took months into days or hours. The feature introduces SageMaker AI model customization agent skills that developers interact with via natural language through popular coding agents such as Kiro, Claude Code, and GitHub Copilot. The capability covers the full lifecycle from use case definition and data preparation through fine-tuning, evaluation, and production deployment to either Amazon Bedrock or SageMaker AI endpoints.

How It Works

• The experience is built on a set of SageMaker AI model customization agent skills that can be installed as a plugin (sagemaker-ai agent plugin) into any compatible IDE, including Visual Studio Code and Cursor, or accessed directly within SageMaker Studio Notebooks where the skills and the Kiro coding agent come pre-installed.
• Developers interact with the agent using natural language prompts; the agent translates intent into executable, editable Python code artifacts that orchestrate the underlying SageMaker infrastructure.
• The skills encapsulate domain expertise across three fine-tuning paradigms: Supervised Fine-Tuning (SFT) for instruction tuning, Direct Preference Optimization (DPO) for tone and preference alignment, and Reinforcement Learning (RL) for tasks with verifiable correctness.
• The agent also handles data format transformation, configures and runs training jobs on managed infrastructure, and applies LLM-as-a-judge metrics for comprehensive quality evaluation.
• Supported model families include Amazon Nova, Meta Llama, Alibaba Qwen, and GPT-OSS variants.
• All generated code artifacts are reusable and editable, enabling integration into AIOps pipelines for reproducibility and automation.

Why It's Important

• Model customization has historically been one of the most technically demanding and time-consuming tasks in the ML lifecycle, requiring deep expertise in distributed training infrastructure, hyperparameter tuning, data formatting standards, and evaluation methodology—barriers that effectively excluded many development teams.
• By abstracting this complexity behind natural language interactions, AWS enables a much broader population of developers, including those without dedicated MLOps or research science backgrounds, to produce production-quality fine-tuned models.
• The reduction from months to days or hours has direct business impact: faster iteration cycles, lower experimentation costs, and quicker time-to-value for AI-powered products.
• The generation of transparent, editable code artifacts also ensures that teams retain full auditability and control, which is critical for regulated industries and enterprise governance requirements.

How It's Different

• Previously, SageMaker fine-tuning required developers to manually write training scripts, configure job parameters via the SageMaker Python SDK or console, manage dataset formatting to framework-specific schemas, and independently set up evaluation pipelines—each step demanding specialized knowledge and significant iteration time.
• While SageMaker JumpStart offered one-click fine-tuning for select models, it provided limited flexibility and no conversational guidance.
• The new agentic experience differs fundamentally by introducing a conversational, intent-driven interface that dynamically generates the full workflow code rather than relying on fixed templates or UI wizards.
• The multi-agent compatibility (Kiro, Claude Code, Copilot) and multi-IDE support distinguish it from point solutions tied to a single toolchain, and the native integration of LLM-as-a-judge evaluation represents a more sophisticated quality assessment approach than traditional metric-only pipelines.

When to Prefer It

• This capability is best suited for teams that need to rapidly prototype and iterate on fine-tuned models without a dedicated ML infrastructure team, or for experienced ML engineers who want to eliminate boilerplate setup and focus on higher-order decisions like architecture selection and evaluation strategy.
• It is particularly valuable when exploring multiple fine-tuning techniques (SFT, DPO, RL) across several model families in parallel, as the agent can scaffold and manage multiple experimental configurations efficiently.
• Teams operating within existing IDE-centric development workflows will benefit from the plugin model, while those already using SageMaker Studio Notebooks gain zero-configuration access.
• Organizations with AIOps maturity will find the reusable code artifact output directly consumable in CI/CD and MLOps pipelines.
• It is less appropriate for highly customized training scenarios requiring bespoke distributed training frameworks or hardware configurations that fall outside the supported model families and techniques.

Availability

• The feature is generally available (GA) as of May 4, 2026, across four AWS regions: US East (N.
• Virginia) us-east-1, US West (Oregon) us-west-2, Europe (Ireland) eu-west-1, and Asia Pacific (Tokyo) ap-northeast-1.
• The Kiro coding agent integration within SageMaker Studio Notebooks requires a separate Kiro subscription.
• Use of Claude Code and GitHub Copilot as alternative agents requires those respective subscriptions and IDE plugin installations.
• No preview or limited-availability caveats are noted in the announcement, though regional expansion beyond the initial four regions has not been announced.
• Customers should consult the SageMaker model customization documentation for the latest information on supported model families, instance types, and any service quota considerations.