A set of six open-source skills now guides coding agents through deploying Hugging Face models to Amazon SageMaker AI. The workflow covers container selection, cloud permissions, endpoint creation, autoscaling, monitoring, smoke tests and teardown, while requiring approval before it creates billable resources.
AWS tested unguided agents from Kiro and Claude Code on recent models and found that both initially selected Text Generation Inference, a once-common server that could not load the chosen Qwen3 architecture. The agents retried versions and generated billable failed deployments before switching to vLLM. On a new multimodal diffusion model, they again produced a TGI-based script even though that server had no suitable backend. The problem was stale deployment knowledge rather than an inability to plan.
The skills move changing facts into editable files that agents load when needed. They discover the AWS account and region, find an existing SageMaker execution role before attempting to create one, resolve a current image from the Deep Learning Containers catalog, attach CloudWatch alarms and run a live endpoint test. Supported targets include real-time, scale-to-zero, serverless, asynchronous and batch inference. The tools use Python and the AWS command line and work with agents that support the skills format. Teams still need suitable permissions, GPU quota and cost controls; a standard real-time endpoint continues billing while it is running.