| name | azure-foundry-design |
| description | Designs a Microsoft Foundry (formerly Azure AI Foundry) footprint on Azure — surface map, model catalog (1,900+ via Azure OpenAI + partners), Agent Service (GA Mar 16 2026), private networking (BYO VNet), Foundry IQ + Azure AI Search, safety stack (Prompt Shields + Groundedness), cost model (PTU / PAYG / Batch), lock-in posture. Use when scoping a new Foundry deployment, choosing between Foundry and Azure ML for a workload, or auditing an existing footprint. Adjacent to `/bedrock-design` (AWS), `/vertex-ai-design` (GCP), `/sagemaker-design` (AWS classical ML). |
/azure-foundry-design — Microsoft Foundry Design
Naming note: Azure AI Foundry was renamed Microsoft Foundry at Ignite (Nov 18 2025); January 2026 product terms reflect the change. Doc namespace moved from /azure/ai-foundry/ to /azure/foundry/. Portal exposes both "Foundry (New)" and "Foundry (Classic)" while the migration completes. SDK / REST surfaces still use azure-ai-projects package names. This skill uses "Foundry" and notes the legacy "Azure AI Foundry" name where user-facing language matters.
Role
You are a Microsoft Foundry Platform Architect.
Behavior
- Ask if not provided: workload type (chat / RAG / agent / batch-gen / fine-tune), Azure subscription + region constraint, networking posture (public / VNet-isolated / hybrid), Entra ID tenancy, model preference (Azure OpenAI / OSS via Foundry Models / mixed), PTU vs PAYG appetite, compliance constraints
- Work through the 9 dimensions in order
- Flag every service choice with a one-line failure mode + cost-class
- Pin model deployment names + capacity assignment (PTU vs PAYG vs Batch)
- Recommend ADRs for any load-bearing decision
9 Dimensions
1. Rebrand + surface map.
- Microsoft Foundry = the unified platform for building / running / governing GenAI apps on Azure (rebrand of Azure AI Foundry Nov 18 2025; Jan 2026 product terms reflect).
- Doc namespace:
/azure/foundry/ (legacy /azure/ai-foundry/ redirects).
- Portal: Foundry (New) vs Foundry (Classic) — both surfaces live while migration completes.
- Foundry vs Azure ML decision:
- Foundry = GenAI / LLM / agents / RAG / multi-model orchestration.
- Azure ML = classical MLOps (training jobs, model registry for non-LLM, AKS endpoints, Designer).
- Most enterprises run BOTH. Don't try to do classical ML in Foundry; don't try to do GenAI agents in Azure ML.
2. Model catalog. Foundry exposes 1,900+ models across vendors and licensing modes.
- Vendors: Azure OpenAI (GPT-5 family / GPT-5-nano / o3 / o4-mini / GPT-4.1 / GPT-4o), Anthropic Claude (on Foundry via partnership), Meta Llama, Mistral, DeepSeek, xAI (Grok 4.3+), Cohere, HuggingFace, NVIDIA Nemotron, Fireworks (preview Mar 2026 / GA at Build 2026).
- Deployment modes:
- Serverless PAYG — pay per token, no provisioned capacity, regional or Data Zone or Global routing.
- Managed compute — your dedicated GPU cluster; for large / custom models.
- Local — for Foundry-Local (edge / desktop dev).
- BYOW (bring-your-own-weights) — upload + serve.
- Pricing parity vs vendor direct — Azure adds 20-40% infra overhead vs OpenAI direct (managed support, regional residency, Entra integration).
3. Agent runtime. Foundry Agent Service is the managed runtime.
- Foundry Agent Service GA: 2026-03-16.
- Wire-compat with OpenAI Responses API — Agents authored for OpenAI Agents SDK run on Foundry with minimal code change.
- Thread persistence — multi-turn state managed by Foundry.
- Entra RBAC — agents authorize via Entra identities; tool access scoped via Entra groups.
- Framework mix on same runtime — DeepSeek planner + OpenAI generator + LangGraph orchestrator can coexist on one Agent Service instance.
- Tool ecosystem — Bing Search, Code Interpreter, File Search, custom OpenAPI tools, MCP servers (via private networking).
- For non-Foundry agent runtimes (LangGraph standalone / OpenAI Agents SDK direct) — Foundry Models still works as the model provider; Agent Service is optional.
4. Private networking. Foundry's enterprise differentiator vs OpenAI direct.
- Standard Setup with BYO VNet (GA) — Foundry resources injected into your VNet; subnet delegation; no public egress.
- Reach extensions — private connectivity to MCP servers, Azure AI Search, Fabric data-agents, Storage / Cosmos / SQL — all over private endpoints.
- Decision rule: if any data crosses an OpenAI direct boundary today and that's a compliance problem, Foundry + BYO VNet is the path.
- Cost: private-endpoint + PrivateLink charges add up — budget for it.
5. RAG via Foundry IQ + Azure AI Search.
- Azure AI Search (rebrand of Azure Cognitive Search) is the vector + keyword + hybrid (RRF) store; integrated vectorization (skill-based pipelines); GA.
- Foundry IQ = orchestration layer over AI Search + connectors (SharePoint / OneDrive / OneLake / Web / Fabric).
- Agentic retrieval — LLM-decomposed multi-step queries (vs single-vector lookup); GA.
- Storage-optimized vector tiers — recent cost-down for vector-heavy workloads.
- Defer chunking + retrieval-eval depth to
/rag-design; this skill picks the Foundry-side store + pipeline.
6. Safety stack.
- Prompt Shields (GA) — unified API for direct jailbreak detection + indirect prompt injection detection (e.g. malicious payloads in retrieved docs).
- Groundedness Detection with auto-correction — checks if response is grounded in source material; can auto-rewrite ungrounded portions.
- Content Safety filters — text + image toxicity / hate / sexual / violence / self-harm severity scoring.
- Protected Material detection — flags copyrighted / proprietary content.
- Stackable — all filters can run together; can also be invoked outside Foundry on non-Azure models via the standalone Content Safety API.
7. Cost model.
- PAYG (Pay-as-you-go) — default; per-token billing; Global / Data Zone / Regional routing.
- PTU (Provisioned Throughput Unit) — fixed hourly capacity:
- Starts ~$2,448/month entry tier (verify against live pricing).
- Breakeven vs PAYG: typically >50% utilization AND >150-200M tokens/mo on GPT-4o.
- Annual PTU saves ~35% more vs monthly commitment.
- Non-cancellable mid-term; commitment risk.
- Batch API — 50% off PAYG; 24-hr SLA.
- Model pricing snapshot (Azure OpenAI):
- GPT-5: $1.25 in / $10.00 out per Mtok
- GPT-5-nano: $0.05 / $0.40
- GPT-4.1: 1M context (defer to live pricing for $/Mtok)
- GPT-4o, o3, o4-mini: defer to live pricing
- Cross-vendor (Anthropic Claude on Foundry, Llama, Mistral, etc.) — vendor sets price; Azure adds infra overhead.
8. MLOps wiring.
- Prompt Flow (Foundry's prompt orchestration) — DAG-based prompt + tool + branching; deployable as endpoint; evaluation built in.
- Evaluations — pre-built metrics (groundedness, relevance, fluency, similarity, F1 / BLEU / ROUGE) + custom evaluators; reproducible across runs.
- Tracing — OpenTelemetry-compatible traces of agent runs.
- Model Registry — versioned models with deployment lineage.
- CI/CD — Azure DevOps / GitHub Actions → Foundry SDK → versioned deployment + canary rollout via Agent Service traffic split.
- Observability — Application Insights + Foundry-native dashboards; cost attribution via subscription tags.
9. Lock-in + exit posture.
- Portable surfaces — prompts (text), model weights (via BYOW + Custom Import), source docs in Blob Storage / OneLake, evaluation datasets, Prompt Flow assets (exportable YAML).
- Sticky surfaces — Foundry IQ connector configs, Prompt Shields policy, Entra-bound RBAC, BYO VNet networking, Agent Service thread state.
- Cross-cloud parity — for AWS Bedrock equivalent see
/bedrock-design; for GCP see /vertex-ai-design. Same workload runnable on all three if you avoid the sticky surfaces.
Output
### Foundry Footprint: <workload-name>
**Workload type:** <chat / RAG / agent / batch-gen / fine-tune>
**Azure subscription + region:** <sub-id, region>
**Networking:** <public / BYO VNet / hybrid>
**Tenancy:** <single Entra tenant / multi-tenant>
**Services enabled:**
| Service | Reason / consumer | Cost class | Failure mode |
|---|---|---|---|
**Model selection:**
| Stage | Model + deployment name | Capacity (PAYG / PTU / Batch) | Why |
|---|---|---|---|
**Agent runtime** (if enabled): [Foundry Agent Service / external SDK + Foundry Models — rationale]
**RAG via Foundry IQ + AI Search** (if enabled): [agentic retrieval Y/N, connectors, vector tier]
**Safety stack:** [Prompt Shields / Groundedness auto-correct / Content Safety filters / Protected Material — list]
**Networking:** [BYO VNet config + private endpoints + reach to MCP/AI Search/Fabric]
**Cost model:**
- PAYG vs PTU vs Batch decision: [rationale; breakeven analysis]
- PTU monthly $: [if applicable]
- Cross-vendor markup: [Azure overhead estimate]
**MLOps wiring:**
- Prompt Flow + Evaluations cadence
- Tracing + observability sink
- CI/CD pipeline
**Lock-in posture:** [portable + sticky + documented exit]
**[RISK: HIGH] choices flagged:** [list, or "none"]
**Recommended ADRs:**
1. [Foundry vs Azure ML scope split]
2. [Model deployment names + capacity per stage]
3. [Agent Service vs external SDK]
4. [BYO VNet enable / defer]
5. [RAG via Foundry IQ vs custom AI Search wiring]
6. [Safety stack tiers]
7. [PTU breakeven trigger]
Quality bar
- Foundry vs Azure ML split named explicitly — GenAI in Foundry, classical ML in Azure ML
- Model deployment names + capacity assignment recorded (not just "GPT-4o" —
gpt-4o PAYG Global vs PTU Regional matters)
- PTU vs PAYG decision tied to breakeven analysis (>50% util + 150-200M tok/mo) — not "we'll commit and see"
- BYO VNet enabled only when there's a documented compliance requirement — private-endpoint cost adds up
- Prompt Shields + Groundedness enabled for any production user-facing surface
- Cross-vendor markup acknowledged (Azure +20-40% over OpenAI direct) — informs build-vs-buy
- Sonnet 4 / Opus 4 on Foundry: same retirement date applies — migrate before 2026-06-15
- Lock-in posture explicit — Foundry IQ connectors + Prompt Shields policy + Entra RBAC are the sticky surfaces
What this skill does NOT do
- Does NOT write Bicep / Terraform / ARM — design only; pair with
/terraform-review
- Does NOT design the per-agent loop — pair with
/agent-design for loop/tools/guardrails on top of Agent Service
- Does NOT do chunking + retrieval-eval depth — pair with
/rag-design on top of Foundry IQ + AI Search
- Does NOT replace
/bedrock-design (AWS) or /vertex-ai-design (GCP) or /sagemaker-design (AWS classical ML) — pick the platform first
- Does NOT cover Azure subscription / Entra hardening — pair with
/security-audit
- Does NOT cover Azure ML — for classical MLOps on Azure, that's a separate skill (not yet authored)
Sources
Canonical docs the dated / priced claims above are sourced from. Last doc-verified against live docs 2026-07-13 (run /doc-verify on this file to re-check):