Each group has entry to the identical basis fashions. The actual aggressive benefit comes from customizing them together with your proprietary information and area experience. However getting there may be complicated, even for knowledgeable groups. It requires mastering fine-tuning methods like Supervised Nice-Tuning (SFT), Direct Choice Optimization (DPO), and Reinforcement Studying Verifiable Rewards (RLVR), navigating fragmented APIs and model-specific information codecs, designing rigorous evaluations, and managing months-long experiment cycles.
Amazon SageMaker AI now gives an agentic expertise that adjustments this. Builders describe their use case utilizing pure language, and the AI coding agent streamlines the complete journey, from use case definition and information preparation by way of method choice, analysis, and deployment. Objective-built agent abilities ship specialised experience on fine-tuning utilized to your particular use case, information transformation to required codecs, high quality analysis utilizing LLM-as-a-Choose metrics, and versatile deployment to Amazon Bedrock or SageMaker AI endpoints. Agent abilities for mannequin customization not solely enhance productiveness but additionally lower token utilization. All generated code is absolutely editable, producing reusable artifacts that combine seamlessly into present workflows.
What makes this expertise really highly effective is agent Abilities for mannequin customization. They’re pre-built, modular instruction units that encode deep AWS and information science experience throughout the complete customization lifecycle. While you describe your use case, the AI coding agent prompts the related abilities, guiding it by way of information preparation and validation, method choice, hyperparameter configuration, mannequin analysis, and deployment. Abilities present specialised information about SageMaker AI APIs, ML workflows, greatest practices, and customary patterns, enabling your coding agent to supply extra correct, SageMaker AI-specific steering, producing ready-to-run notebooks at every step. Abilities are absolutely customizable, so you’ll be able to modify them to match your crew’s workflows, governance requirements, and tooling preferences, enabling reproducible organizational greatest practices, a typical problem with general-purpose coding assistants.
Amazon Kiro in SageMaker AI Studio JupyterLab
JupyterLab in SageMaker AI contains an built-in agentic growth surroundings help by way of ACP. By default, Kiro, Amazon’s AI software program growth agent, is pre-configured within the chat panel, offering AI-powered code completion, debugging help, and interactive coding help straight inside your JupyterLab surroundings. While you use coding brokers in SageMaker AI JupyterLab, the house robotically hundreds related SageMaker AI mannequin customization Abilities into your agent’s context.
Moreover, you’ll be able to configure different Agent Communication Protocol (ACP) appropriate coding brokers of your selection, comparable to Claude Code, providing you with flexibility to work with the instruments that greatest suit your workflow. ACP-compatible brokers can profit from the identical SageMaker AI Abilities integration when used inside SageMaker AI JupyterLab. Whereas this instance reveals the combination with JupyterLab, you may as well use distant entry to your individual IDE outdoors of JupyterLab.
Conditions
Earlier than beginning this tutorial, you have to have the next conditions:
- An AWS account
- The power to entry or create a SageMaker AI area. In the event you don’t have a SageMaker AI area, you’ll be able to create one utilizing the fast setup or guide setup choices
- An AWS IAM function with the required permissions
- An Amazon Easy Storage Service (Amazon S3) bucket
- Entry to or can create a SageMakerAI Studio JupyterLab compute house. There is no such thing as a minimal occasion kind requirement to make use of the brand new options.
- As of this publication, SageMaker AI Distribution picture model 4.1 or greater is required in your SageMakerAI Studio JupyterLab.
- Confirm or Connect AmazonSageMakerFullAccess managed coverage to your area’s execution function. Connect the extra inline coverage for Lambda, S3 and Bedrock entry to the identical execution function
- Your SageMakerAI Studio execution function’s belief coverage should permit these three companies to imagine the function: sagemaker.amazonaws.com, lambda.amazonaws.com, bedrock.amazonaws.com.
Abilities overview
The SageMaker AI agent abilities are constructed conforming with the Agent Abilities open format. The agent-guided mannequin customization workflows are powered by 9 modular abilities that cowl the complete customization lifecycle:
Ability Identify
Part
Description
Use Case Specification
Configuration
Structured discovery to outline enterprise drawback, customers, and success standards
Planning
Discovery
Generates a dynamic, multi-step customization plan tailor-made to your use case
Nice-tuning Setup
Configuration, Coaching
Selects base mannequin from SageMaker AI Hub and recommends method (SFT, DPO, or RLVR)
Dataset Analysis
Analysis, Coaching
Validates dataset format and schema earlier than coaching
Dataset Transformation
Information Engineering
Converts between ML information codecs (OpenAI chat, SageMaker AI, Hugging Face, Amazon Nova)
Nice-tuning
Coaching
Generates coaching notebooks for SageMaker AI serverless fine-tuning
Mannequin Analysis
Analysis
Configures LLM-as-Choose analysis with built-in and customized metrics
Mannequin Deployment
Deployment
Determines deployment pathway (SageMaker AI endpoint or Bedrock) and generates code
The coding agent (Kiro, Claude Code, Cursor, and so on.) offers the conversational interface whereas the SageMaker AI Abilities orchestrate the workflow. While you work together together with your coding agent, it prompts the related abilities. This lets you name SageMaker AI APIs, entry S3 information sources, and work together with mannequin registries by way of AWS-provided MCP servers. Jupyter notebooks are generated for you that execute every step of the method into present ML pipelines.
Supported Nice-Tuning Methods
The mannequin customization abilities at the moment help three fine-tuning methods and advocate the proper one in the course of the planning section based mostly in your use case.
Approach
Description
Finest For
SFT (Supervised Nice-Tuning)
Trains on enter/output pairs
Job-specific habits: instruction following, format compliance, domain-adapted responses
DPO (Direct Choice Optimization)
Trains on most popular vs. rejected outputs
Aligning tone, model, and subjective preferences to match human judgement
RLVR (Reinforcement Studying with Verifiable Rewards)
Trains utilizing code-based reward features
Duties the place correctness might be programmatically verified
Resolution implementation
For this resolution, you’ll fine-tune a small language mannequin (SLM) on the FreedomIntelligence/medical-o1-reasoning-SFT dataset to construct a scientific reasoning mannequin that walks by way of medical instances step-by-step earlier than offering a analysis. This demonstrates how fine-tuning can specialize a general-purpose mannequin for domain-specific reasoning duties. In the event you’d prefer to attempt a distinct use case, SageMaker AI offers a library of pattern datasets throughout methods like SFT, DPO, and RLVR that you need to use as a place to begin.
Getting began
- Open or Create a SageMaker AI House with JupyterLab
- Navigate to SageMaker AI Studio
- Go to Areas within the left navigation panel or click on “Customise with agent” from the mannequin hub
- Both:
- Click on Create House and choose JupyterLab as your software
- Open an present House that features JupyterLab
On this put up, we’ll begin with utilizing Kiro and swap to Claude Code as our coding agent. To maintain utilizing Kiro, transfer to the Planning Part part, or transfer to the subsequent part to see learn how to use Claude Code in JupyterLab.
Begin Utilizing Kiro within the Chat Panel:
Kiro requires authentication earlier than you need to use it. The chat panel will information you thru the authentication course of.
- In JupyterLab, open the chat panel by clicking the chat icon in the proper sidebar
- Sort @ to see your out there brokers
- Choose @Kiro from the agent dropdown. Begin asking questions or requesting code help.
The primary time you utilize Kiro in an area, it’s going to ask you to login. To login, comply with the directions offered by the chat, or comply with right here:
- In JupyterLab, open a brand new terminal: File > New > Terminal
- Run the next command kiro-cli login –use-device-flow
Choose one of many 3 login choices within the terminal:
- Use for Free with Builder ID
- Use for Free with Google or GitHub
- Use with Professional license
- Enter a immediate: “I need to customise a mannequin”
Configuring Claude Code in JupyterLab
SageMaker AI Studio helps implementing further coding brokers utilizing Agent Management Protocol (ACP). Instance brokers that help ACP embrace:
- Claude (through claude-agent-acp)
- OpenCode (through opencode CLI >= 1.0.0)
- Gemini (through gemini CLI >= 0.34.0)
- Codex (through codex-acp)
View the JupyterLab person information for extra data on set up steps.
To make use of Claude Code:
- Set up the CLI device in your SageMaker AI Studio JupyterLab terminal:
npm set up -g @zed-industries/claude-agent-acp - Restart the house by operating the command restart-jupyter-server or by restarting the house through the Studio UI. Please notice, it will end in any unsaved work or in reminiscence state (like lively kernels) being misplaced.
- Authenticate with the agent following its particular authentication course of
- Choose the agent from the persona dropdown within the JupyterLab chat panel (@Claude)
Claude Code can be utilized with most Anthropic subscriptions together with configuring Claude Code with Amazon Bedrock on Amazon SageMaker AI Studio. To configure Claude Code to make use of Claude by way of Amazon Bedrock comply with the conditions within the Claude code information, enabling Bedrock mannequin entry and offering your execution function entry to bedrock:InvokeModel and bedrock:InvokeModelWithResponseStream. Then, create the next file to configure Claude Code to make use of Bedrock.
~/.claude/settings.json:
{
“env”: {
“CLAUDE_CODE_USE_BEDROCK”: “1”
}
}
Planning section
Upon receiving the person immediate, the coding agent doesn’t instantly start executing duties. It enters a planning section the place it identifies and prompts the talents essential to finish the job. Within the course of, the agent generates a workflow which customers can overview and modify. From the preliminary immediate, the agent acknowledges two related talent domains and prompts each the planning talent for structuring the general workflow and the finetuning-setup talent for configuring the coaching job. Earlier than producing any code, the agent asks focused questions on dataset readiness and use case particulars to tell its method and analysis metrics suggestions.
Nice-tune in SageMaker AI
With a number of mannequin households and fine-tuning methods out there, selecting the best strategy in your particular use case might be difficult. The agent analyzes your dataset construction and job necessities to supply tailor-made mannequin and method suggestions, serving to you keep away from expensive trial-and-error cycles. SageMaker AI helps serverless customization throughout Amazon Nova, GPT-OSS, Llama, Qwen, and DeepSeek household of fashions. For this use case, we selected Qwen3-0.6B as a result of it’s cost-effective to coach and deploy whereas being ample for domain-specific duties like medical reasoning.
- Within the chat panel, immediate the agent: “I need to fine-tune a mannequin for scientific reasoning that walks by way of medical instances step-by-step earlier than offering a analysis.”
- Verify the plan and reply the agent’s follow-up questions. The agent generates a coaching pocket book that can use a SageMaker AI serverless coaching job with coaching and validation metrics tracked by way of built-in SageMakerAI MLflow Apps.
- Open the pocket book, confirm the code and run the pocket book cells to submit the coaching job.
- Monitor the job inside your SageMaker AI Studio.
The mannequin’s loss will present a gradual lower throughout coaching, exhibiting it efficiently discovered to supply step-by-step scientific reasoning earlier than reaching diagnoses. For a deeper take a look at the complete metric set and per-step breakdowns, we will view extra within the MLflow app.
Analysis
As soon as coaching completes, we have to measure how nicely the fine-tuned mannequin performs in comparison with the bottom mannequin. The agent recommends an analysis strategy based mostly on our use case, or we will specify the metrics we care about, comparable to accuracy on held-out medical reasoning questions or reward rating enchancment over the bottom mannequin. It then generates a pocket book in SageMaker AI Studio JupyterLab that runs the analysis towards an analysis dataset and studies the outcomes, so we will validate the mannequin’s efficiency. Analysis outcomes are additionally distributed to MLflow for comparisons earlier than shifting to deployment.
Deployment
With analysis full, the ultimate step is deploying the fine-tuned mannequin for inference. The agent walks us by way of deployment choices throughout SageMaker AI and Bedrock by way of Bedrock Customized Mannequin Import, relying on our latency, scaling, and integration necessities. It then generates a pocket book in JupyterLab that provisions the endpoint and runs a pattern inference request, so we will validate whether or not the deployed mannequin is able to serve predictions.
Customise abilities
The abilities included with SageMaker AI cowl widespread fine-tuning workflows, however you may as well customise present abilities or creator new ones to match your group’s requirements and tooling. For instance, you would possibly prolong the model-evaluation talent to incorporate domain-specific metrics or add a brand new talent for a customized deployment goal. Abilities are outlined in easy markdown information within the ~/.kiro/abilities listing, making them simple to creator, version-control, and share throughout your group.
Conclusion
On this put up, we walked by way of the mannequin customization lifecycle utilizing SageMaker AI agent abilities. Ranging from a single pure language immediate, the agent deliberate the workflow, configured and ran a SFT fine-tuning job on Qwen3-0.6B, evaluated the outcomes with metrics tailor-made to our use case, and deployed the fine-tuned mannequin. The agentic mannequin customization expertise in Amazon SageMaker AI is on the market at the moment. You may get began in minutes. Merely launch a JupyterLab house in SageMaker Studio with Kiro and Agent Abilities pre-configured, or carry the identical Abilities into your most popular IDE from GitHub. Describe your use case in pure language, and let the agent information you from information preparation by way of analysis and deployment.
What as soon as required months of specialised ML work and deep information can now be accomplished in days. The experience is encoded. The workflow is guided. And the code is yours. Get began at the moment by visiting the GitHub repository for the SageMaker AI agent abilities plugin and step-by-step information. Evaluation the documentation to see how SageMaker AI serverless mannequin customization with agent abilities can speed up your path from thought to manufacturing fashions.
Concerning the authors
Lauren Mullennex
Lauren is a Senior GenAI Specialist Options Architect at AWS. She has over a decade of expertise in ML, DevOps, and infrastructure. She is a broadcast creator of a ebook on pc imaginative and prescient. Outdoors of labor, you could find her touring and mountaineering along with her two canine.
Sandeep Raveesh
Sandeep is a Generative AI Specialist Options Architect at AWS. He works with buyer by way of their AIOps journey throughout mannequin coaching, generative AI purposes like brokers, and scaling generative AI use-cases. He additionally focuses on go-to-market methods serving to AWS construct and align merchandise to resolve business challenges within the generative AI house. You’ll be able to join with Sandeep on LinkedIn to find out about generative AI options.
Mike Diamond
Mike is a Principal Product Supervisor for Amazon SageMaker AI. With twenty years of expertise making use of AI to high-stakes domains, Mike is captivated with accountable AI and making machine studying extra accessible by way of agentic workflows and developer-friendly tooling.
Joshua Towner
Joshua Towner is a Senior Software program Engineer at AWS.
Bobby Lindsey
Bobby Lindsey is a Machine Studying Specialist at Amazon Net Providers. He’s been in expertise for over a decade, spanning varied applied sciences and a number of roles. He’s at the moment targeted on combining his background in software program engineering, DevOps, and machine studying to assist prospects ship machine studying workflows at scale. In his spare time, he enjoys studying, analysis, mountaineering, biking, and path operating.
Emily Moeng
Emily Moeng is a Language Information Science Supervisor at AWS with a background in theoretical and experimental linguistics. She makes a speciality of distilling AI/ML targets into sturdy, execution-ready pipelines for information curation, annotation, and mannequin analysis.
Vineet Sharma
Vineet is a Senior Product Advertising and marketing Supervisor, Tech at AWS, targeted on Amazon SageMaker AI. He makes a speciality of go-to-market technique, product launches, and translating complicated AI and ML companies into compelling buyer worth. He’s captivated with creating nice buyer experiences by way of clear, impactful messaging. Join with him on LinkedIn.
