Microsoft Foundry Ep5: Architecture Agent, Cost Agent and Human Sign-off

An AI agent can draw a design and price it in a minute, but who says yes to the money? In Ep5 of my Microsoft Foundry series, I build an architecture agent that picks from an approved catalog and a cost agent that compares three options from a price sheet. Then the architect and finance sign off, and we read the real Azure bill. For IT and security pros putting AI agents near design and budget calls. What this would cost you 1) Real Azure cost on screen: $3.35 actual Oct 2026 spend ($6.55 forecast) under a $10/mo budget in Cost analysis (AccumulatedCosts, ungrouped, read Oct 23, 2026). Billing lags up to a day; any delta goes in the pinned comment. 2) Licenses: none needed. Prompt agents without Code Interpreter or File Search have no separate fee (source 4); patterns.json, prices.json and gates.json are text files (prices.json uses fictional prices for local math). 3) Estimated total for you, model tokens only: about $0.45 (under $0.70 ESTIMATE) on gpt-5.1, or under $0.15 (ESTIMATE) on gpt-5-mini (not Azure's figure). Basis: 4 paid calls (~68K tokens on UI counters; ~28K in, 40K out incl. reasoning assumed), East US 2 Global Standard read Oct 24, 2026 (source 7): gpt-5.1 $1.25 in, $10.00 out/1M; gpt-5-mini $0.25 in, $2.00 out/1M. 10K TPM; evaluations off ($0.02 in, $0.06 out/1K, source 2). Level: Intermediate Narration: my own voice. Some illustrations are AI-generated. Ep5 of AI Agents the Safe Way. Start here (Ep0):    • Microsoft Foundry: AI Agents With Humans i...   Ep3:    • Microsoft Foundry Ep3: Build a Vetting Age...   | Ep4:    • Microsoft Foundry Ep4: A Security Agent, a...   What I set up Deployment (Step 1): Global Standard, 10,000 tokens per minute (10K TPM), DefaultV2, auto-upgrades off, Succeeded. Narration mentions gpt-5-mini; recording runs on gpt-5.1 (noted on screen). Inputs (Step 2): patterns.json (3 patterns; 3 placement rules: keep ticket DB, no new primary DB or sign-in, firewall via change board) and prices.json (fictional prices so every number traces to a line). Architecture agent (Steps 3-5): built by demo user with Foundry User only (TTT-Architecture-51 on screen, tools none, knowledge none). Picks PAT-WEB-READVIEW-01 for Case 1 (Sensitive), stops at READY_FOR_DESIGN, and refuses an ask to add a new DB by citing its placement rule. Cost agent (Step 6): TTT-Cost (tools none, knowledge none) reads architecture.md and prices.json and returns 3 options (A: $24,000 build / $650/mo, recommended; B: $28,800 / $950/mo; C: $31,200 / $2,450/mo), an unsigned $40,000 envelope, and assumptions from the made-up sheet. Sign-off #3 and real bill (Steps 7-8): two GATE-3 records (architect + finance, fictional names, entered by script for demo; no agent writes gates.json) move status to DESIGNED. Cost analysis shows $3.35 actual, $6.55 forecast ($10/mo budget). Lab stays up for the series; one cleanup at the end. What we know Foundry User builds and tests agents without deploy or RBAC rights (source 3). Global Standard bills per token; TPM limits cap spend rate (sources 5, 6). Reasoning tokens bill as output (source 5); evaluations bill when on (source 2). Foundry retires workflows Dec 1, 2026 for Agent Framework (source 1). What we don't know, or don't cover Your bill depends on model, region and usage. One refusal isn't proof an agent obeys every time; gates.json stays separate. Not covered: firewall rules (Ep6), live pricing feeds, production sign-in. Chapters 00:00 Two agents draw and price it 00:53 Series recap and lab setup 01:34 Pattern, TCO and money envelope 02:33 Step 1: Check the model deployment 03:09 Step 2: Catalog and made-up price sheet 03:57 Step 3: Build TTT-Architecture 04:44 Step 4: Run the design for Case 1 05:17 Step 5: Can it add a new database? 05:43 Step 6: Build and run TTT-Cost 06:32 Step 7: Sign-off #3: architect + finance 07:08 Proof: the real Azure bill 07:45 Six mistakes to skip 08:34 Recap table and next episode Jargon Pattern: pre-approved design shape. TCO: build cost + monthly run rate. Run rate: monthly cost once live. Money envelope: max budget, signed by finance. Sources (checked Oct 24, 2026; 1-6 Microsoft Learn, 7 Azure) 1. Workflows. https://learn.microsoft.com/azure/fou... 2. Observability. https://learn.microsoft.com/azure/fou... 3. Foundry RBAC. https://learn.microsoft.com/azure/fou... 4. Agent FAQ. https://learn.microsoft.com/azure/fou... 5. Reasoning. https://learn.microsoft.com/azure/fou... 6. Deployments. https://learn.microsoft.com/azure/fou... 7. Azure Retail Prices API, East US 2. https://prices.azure.com/api/retail/p... Corrections: none as of October 24, 2026 Teach-only: per Microsoft docs. Fictional company, people and price sheet; screen actions recorded by the team from Jim's script.