agentic-deploy

Agent Skills that let a coding agent (Claude Code, Codex) deploy models to Amazon SageMaker safely and repeatably — picking the right serving container, ensuring an execution role exists, and creating endpoints with autoscaling, alarms, and tagging on by default.

The skills wrap boto3 and AWS's published Deep Learning Containers catalog directly; the bundled scripts don't require the SageMaker Python SDK, though using the SDK alongside is fine — see python-env-setup/SKILL.md for the trade-offs and the few SDK cases that still need care.

Skills

Live in agentic-deploy-skills/sagemaker-skills/. sagemaker-deployment-planner is the entry point; it coordinates the rest:

Skill	Role
sagemaker-deployment-planner	Entry point — asks clarifying questions, picks a pathway, coordinates the others
aws-context-discovery	Reads local AWS profile, region, account, caller identity
python-env-setup	Isolated Python env (3.10–3.12) with current boto3/awscli
serving-image-selection	Picks the serving container (vLLM, TEI, etc.) and resolves its image URI
sagemaker-iam-preflight	Finds or creates a usable SageMaker execution role
sagemaker-production-defaults	Creates the endpoint (real-time or async) with autoscaling, alarms, tagging

Each skill is a directory with a SKILL.md plus bundled scripts/ and references/.

Install

Symlink the skills into your agent's skills directory (auto-detects Claude Code and Codex):

cd agentic-deploy-skills
bash install.sh                 # all detected agents; symlinks → `git pull` keeps them current
bash install.sh --claude        # Claude Code only (~/.claude/skills)
bash install.sh --copy          # standalone copies instead of symlinks
bash install.sh --uninstall     # remove what it installed

Idempotent, and won't clobber unrelated entries without --force. See bash install.sh --help.

Deployment flow

deployment-planner → aws-context-discovery → python-env-setup
   → serving-image-selection → iam-preflight → production-defaults

The production deploy scripts (deploy.py, deploy_async.py) consume the values the earlier skills produce (region, image URI, role ARN) and emit a machine-readable JSON summary. Tear down with teardown.py <endpoint-name> <region>.

Validation

Structural CI checks every skill is well-formed (valid frontmatter, scripts parse, referenced files exist). Runs on Python 3.12, no AWS/network needed:

cd agentic-deploy-skills
pip install pyyaml
python scripts/validate_skills.py sagemaker-skills

See agentic-deploy-skills/scripts/README.md for what it does and doesn't cover.

Layout

agentic-deploy-skills/      # the skill bundle (the deliverable)
  install.sh                # symlink/copy skills into an agent's skills dir
  sagemaker-skills/         # one directory per skill
  scripts/validate_skills.py

Agent eval runs (claude_code_tests/, codex_tests/, kiro_test/) are kept local and gitignored — they're not part of the published skills.