Lock the approved stack

Owner: CoE Lead + Platform Team. Input: Operating model in place. Sub-steps:

1. Pick (and write down) the approved stack (framework.md §16):
   • LLM(s): primary + fallback. Multi-vendor from day one.
   • Embeddings model
   • Orchestration: one visual (e.g., n8n / Power Automate) + one code-first (e.g., LangGraph). Pick the one(s) you'll actually use.
   • Vector store
   • Observability platform (LangSmith / Helicone / Arize / etc.)
   • Identity provider — use the existing one.
   • Secret store — use the existing one.
   • Source control — use the existing one.
   • Registry / intake tool (Notion / Airtable / ServiceNow / Jira — pick one).
   • Open protocols: MCP default for tool integration where supported.
2. Set up dev / test / prod environments (separate at minimum logically; ideally physically).
3. Publish the approved stack list internally. Output / gate criteria: A 1-page "Approved AI Stack" document. Dev / test / prod environments exist. Decision branches:

• Multiple departments push for different orchestrators → resolve before any agent is built. Exceptions are allowed; surprise exceptions are not. Skip-this-step risk: Tool sprawl. Each department buys its own platform; standards become unenforceable.