Amazon Launches aws-ai-ml Skill for SageMaker Inference Optimization via MCP Coding Agents
Summary
- • Amazon's aws-ai-ml skill turns MCP coding agents into SageMaker inference optimization experts
- • Compatible with Kiro, Claude Code, and Codex; setup takes under 10 minutes
- • Agents benchmark endpoints, find optimal instances, and generate SageMaker SDK v3 code
- • Available locally via Agent Toolkit for AWS or in SageMaker Studio JupyterLab
Details
aws-ai-ml skill released via Agent Toolkit for AWS
Amazon launched the aws-ai-ml skill on October 5, 2026, available through the Agent Toolkit for AWS. It plugs into the Model Context Protocol layer, turning any MCP-compatible coding agent into a SageMaker AI inference optimization expert.
MCP integration works with Kiro, Claude Code, and Codex
Compatible agents include Amazon Kiro, Anthropic Claude Code, and OpenAI Codex. Kiro and Claude Code can discover and load the skill at runtime through the AWS MCP Server without local installation.
Natural-language prompts yield benchmark reports and ranked deployment options
Engineers describe their intent — a cost envelope, performance target, or model to evaluate — and the agent benchmarks live endpoints, recommends instance types with throughput and latency metrics, compares runs, and generates executable Python code. It asks clarifying questions rather than guessing.
Qwen3-8B showed 44–47% higher throughput vs Qwen3-1.7B in benchmark example
A benchmark comparison in the release showed Qwen3-8B on a 4-GPU ml.g5.12xlarge delivered ~44–47% higher output token throughput and lower request latency than Qwen3-1.7B on an ml.g6.4xlarge, though Qwen3-1.7B had lower time-to-first-token on a single GPU.
Two installation paths: local Agent Toolkit or SageMaker Studio JupyterLab
Local install requires AWS CLI 2.35+ and uv; the command 'npx skills add aws/agent-toolkit-for-aws/skills/aws-ai-ml' adds the skill in minutes. Alternatively, a pre-configured SageMaker Studio JupyterLab space ships with the skill built into the image.
Agent confirms before impactful actions and outputs auditable code only
Safety guardrails require explicit user confirmation before benchmarking live endpoints (which drives real traffic). Every agent output is expressed as human-readable, modifiable SageMaker Python SDK v3 code — no opaque UI operations.
Rows cover the skill's capabilities, compatibility, performance benchmarks, deployment options, and safety guardrails.
What This Means
Amazon is making it dramatically easier for engineers to optimize AI model deployments on SageMaker by embedding inference expertise directly into the coding agents they already use. Rather than studying SageMaker's broad infrastructure surface area, engineers can simply describe their goal in natural language and receive grounded, executable configuration code. This reflects a broader industry trend of packaging cloud expertise as MCP 'skills' — modular plug-ins that make agentic workflows more capable without retraining. For teams already using Claude Code, Codex, or Kiro, this effectively adds a senior solutions architect to their AI assistant at no extra setup cost.