Run Claude Code in Production Without Building a Runtime Team

A practical architecture for turning Claude Code workflows into isolated, logged, API-driven production runs.

Audience: AI infra founders and Staff Engineers moving a Claude Code prototype into a customer-facing workflow.

The problem

A local Claude Code demo usually depends on one laptop, ambient credentials, loose files, and human supervision. Production needs repeatable inputs, isolated execution, logs, artifacts, and a stable API response.

Implementation path

Package the workflow as a SKILL.md folder, upload it to Argo, pass customer files at run time, execute inside a scoped sandbox, keep provider credentials in the broker, and return a typed result with downloadable artifacts.

Tradeoffs and failure modes

You give up raw machine control in exchange for a safer boundary, durable logs, and a result contract your product can call repeatedly.

Production run contract

Input files -> Argo run -> isolated sandbox -> Claude Code tool calls -> argo.result.v1 -> signed artifacts

curl -X POST "$ARGO_API_URL/api/skills/<skill_id>/run" \
  -H "Authorization: Bearer $ARGO_API_TOKEN" \
  -F "prompt=Run the playbook against the attached inputs" \
  -F "documents=@./sample.pdf"

Run this on Argo