AWS introduced an MCP-compatible skill that helps coding agents benchmark SageMaker endpoints, compare deployment options, and generate Python SDK v3 code.
Amazon SageMaker AI introduced the `aws-ai-ml` skill on October 5, 2026, giving compatible coding agents tools for generative AI inference optimization. Available through the Agent Toolkit for AWS, the skill works with agents including Kiro, Claude Code, and Codex, as well as other agents that support the Model Context Protocol.
A developer can describe a goal—such as measuring an existing endpoint or finding a suitable instance type—and the agent can produce executable SageMaker Python SDK v3 code. The workflow can benchmark endpoints, recommend deployment configurations, and compare previous benchmark runs. AWS says outputs are inspectable code that users can modify and run in their own environment.
Benchmark reports can include throughput, latency percentiles, time to first token, and concurrency. AWS’s example compares two Qwen3 models using a 512/256-token workload at concurrency four. The Qwen3-8B configuration recorded 271.2 output tokens per second, compared with 188.2 for Qwen3-1.7B. However, the setups used different hardware: the larger model used four A10G GPUs, while the smaller one used a single L4 GPU. AWS attributes much of the throughput difference to the extra compute, so the comparison does not isolate model size as the cause.
The agent is designed to ask for missing details rather than guess, and AWS says it requests confirmation before load-testing a live endpoint. Users need AWS credentials with permissions for SageMaker operations such as creating endpoints and running benchmark or recommendation jobs. The skill can be installed for a local coding agent or used in a preconfigured SageMaker Studio JupyterLab space.