How to use the SageMaker Production Defaults agent skill

What the skill makes the agent do

An agent skill is a folder of instructions and scripts that an AI agent loads when a task matches the skill's description. This one applies when an agent is about to create a SageMaker endpoint, which is a hosted server that runs an AI model. Instead of writing a bare create call, the agent runs ready-made Python scripts.

The scripts create the model, endpoint config, endpoint, autoscaling target and policy, and CloudWatch alarms as one unit. Every resource gets a consistent tag set that includes CreatedBy=agentic-deploy-skills so it can be found later for cleanup. Autoscaling is set to target tracking on invocations per instance, and data capture to S3 stays off unless you turn it on.

• deploy.py: real-time endpoints
• deploy_ic.py: real-time endpoints that scale to zero instances using inference components
• deploy_async.py: async endpoints, which also scale to zero
• invoke_endpoint.py: sends a test request to a real-time endpoint
• teardown.py: removes resources created by the skill

How to install the skill

Download the skill ZIP from this page and unzip it. For a personal install, place the skill folder in ~/.claude/skills/. For a single project, place it in that project's .claude/skills/ folder. The folder must include SKILL.md and the scripts/ directory.

In the Claude apps you can instead upload the ZIP under Settings → Capabilities → Skills. If an installed copy has no scripts/ directory, the skill notes say to fetch the scripts from the source repository rather than rewriting them from the description.

How to trigger the skill with natural requests

You do not need to name the skill. The agent loads it when you ask for a SageMaker endpoint deployment, or when it is about to write code that calls create_endpoint. Good requests state the model, the container image, the IAM role and the region, or say that those are already chosen by other steps in your workflow.

The skill is meant to be the final step in a deployment workflow. Image selection, IAM checks and region discovery should happen first, because the skill expects their results as inputs.

• Deploy the model at s3://your-bucket/qwen3-medical/model.tar.gz as a SageMaker real-time endpoint on ml.g5.xlarge with autoscaling and alarms.
• Create an async SageMaker endpoint that scales to zero for my batch jobs.
• Set up a real-time endpoint for bge-large-en-v1.5 on ml.c6i.2xlarge in my region.

Run a deployment step by step

For a text-generation model served with vLLM, the skill expects a model.tar.gz artifact that you have packaged and uploaded to a bucket you control. SageMaker extracts it to /opt/ml/model, so the endpoint does not fetch repository code at runtime. The example below uses placeholder values for the image URI, role ARN, bucket and region.

The script creates resources in order, waits up to 30 minutes for InService, reports failure reasons, registers autoscaling and alarms, and prints the teardown command. It writes a JSON summary with the endpoint, config and model names to stdout for later scripts.

• Image URI and AMI version come from the image-selection step.
• Role ARN comes from the IAM preflight check.
• Region comes from the AWS context discovery step.
• TEI embedding models do not need the --inference-ami-version flag, which is vLLM-specific.

Test the endpoint before you trust it

InService only means the container answered its ping check. In MMS-based containers the front end can answer pings while the Python worker crash-loops, so the skill requires two checks before it reports success. First, send one real request. For real-time endpoints, use invoke_endpoint.py and require an HTTP 200 with a sensible body. For async endpoints, upload one input to S3, invoke the endpoint, and poll the output location.

Second, scan the endpoint logs for worker-crash messages such as Worker died, Load model failed or ImportError. Inference-component deployments log to a different group, and deploy_ic.py scans that group automatically while it waits.

• If the log scan finds a match, the skill is meant to show the actual traceback from CloudWatch rather than the InService status.
• If a diagnostic call is denied, for example logs:FilterLogEvents, the skill says so in one line and carries on with the checks it can run.

Safety notes and tips

The scripts create billable AWS resources. Endpoints, CloudWatch alarms and optional data capture to S3 all incur ongoing charges, so run teardown.py when you are finished. Only run the skill in an AWS account where you have permission to create SageMaker, IAM, CloudWatch, Application Auto Scaling and tagging resources.

The skill runs scripts and uses the network to call AWS APIs. Review the scripts before running them in a shared account. Keep SM_VLLM_TRUST_REMOTE_CODE=false unless a specific model architecture requires custom code you have reviewed. Loading weights from the Hub at runtime is an explicit exception and can take 5 to 15 or more minutes after startup.

• Use a Hugging Face token only for gated models loaded from the Hub.
• Keep the minimum capacity at 1 for a warm endpoint, or accept cold-start delay if you need scale to zero.
• Enable data capture only when you need request logs, because it adds S3 storage costs.