Microsoft Foundry Ep3: Build a Vetting Agent, Then a Person Signs Off

Before anyone builds anything, someone with a budget has to say yes. I build a vetting agent in Microsoft Foundry that checks a request for gaps and duplicates and scores it, but can't approve it. A person, the sponsor, makes the call and writes it down. For IT and security pros whose teams want AI to triage requests. What this would cost you 1) Azure's real total for this recording day: pending, added within about a day in the pinned comment (Azure's billing data lags by up to a day). 2) Licenses: none needed. Foundry Agent Service has no separate charge for prompt agents without Code Interpreter or File Search; you need an Azure subscription and roles (source 4). 3) Estimated total for you, model tokens only: under $0.10 (ESTIMATE, not Azure's figure). Basis (call count and tokens assumed, not measured): about 4 agent calls, about 5K input and up to 10K output tokens each incl. reasoning, gpt-5-mini Global Standard, East US 2, $0.25 in, $2.00 out per 1M, read Oct 16, 2026 (source 7). Playground evaluations off ($0.02 in, $0.06 out per 1K tokens, source 7). Level: Intermediate Narration: my own voice. Some illustrations are AI-generated. This is Ep3 of my Microsoft Foundry series, AI Agents the Safe Way. New to AI agents? Start here (Ep0):    • Microsoft Foundry: AI Agents With Humans i...   Ep2 (intake agent and case 1):    • Microsoft Foundry Ep2: Build the Intake Ag...   What I set up The model back: gpt-5-mini as ttt-model, Global Standard, 10,000 tokens per minute, automatic upgrades off, default guardrail. Kept for Ep4; deleted after Ep4. Playground evaluations switched off before the first chat: metrics, unselect every evaluator. TTT-Vetting, a prompt agent built by a demo user with only Foundry User. Tools none, knowledge none, a written rubric, three made-up earlier cases, a never list and a JSON schema whose only statuses are submitted and ready for scope. Case 1 from Ep2 run through it once: a vetting brief with a score against the rubric. Then an ordinary request to mark it scoped: it says no, and the status stays ready for scope. Shown as returned. Sign-off #1: the sponsor's decision typed by hand into gates.json. No agent writes that file. What we know (Microsoft docs) Evaluations in the agents playground are on by default for Foundry projects and are billed; you turn them off under metrics (source 1). An evaluation uses a model as a judge to score agent answers (source 2). gpt-5-mini is a reasoning model; reasoning tokens are billed as output tokens. It needs no access request (source 5). Foundry User can build and test agents but can't deploy models or complete role assignments (source 3). Global Standard bills per token processed (source 6). What we don't know, or this video doesn't cover Your bill: it depends on model, region and usage. Mine is an estimate plus Azure's real total for one day. Which meter playground evaluations bill to; Microsoft's docs don't say. Whether the agent follows its rules every time. One run isn't proof; rerun the test when instructions change. Not covered: search indexes, tools, the security agent (next episode). Chapters 00:00 Agents score, a person decides: what we build 00:25 Where we are in the series 01:19 Put the model back: gpt-5-mini, Global Standard 02:09 The meter that's on by default: playground evaluations 03:04 Vetting's job, and where it has to stop 03:51 Gaps, the score and the duplicate check 04:56 Build the TTT-Vetting agent 06:09 Vetting case 1 07:05 Can it approve itself? It says no 07:52 Sign-off #1: the sponsor decides 08:58 Sign-off checklist before real requests 09:43 Before real requests: what stays running 10:07 Cost note, next episode and 3 takeaways Jargon Vetting agent: an AI agent that checks a request for gaps and duplicates and scores it. Rubric: a written list of what earns points. Sponsor: the person who owns the problem and the budget. Sources (Microsoft Learn, checked Oct 16, 2026; 7: Azure) 1. Observability in generative AI, Pricing. https://learn.microsoft.com/azure/fou... 2. Agent evaluators. https://learn.microsoft.com/azure/fou... 3. Role-based access control for Microsoft Foundry. https://learn.microsoft.com/azure/fou... 4. Foundry Agent Service FAQ. https://learn.microsoft.com/azure/fou... 5. Azure OpenAI reasoning models. https://learn.microsoft.com/azure/fou... 6. Deployment types in Microsoft Foundry Models. https://learn.microsoft.com/azure/fou... 7. Azure Retail Prices API, East US 2, read Oct 16, 2026. https://prices.azure.com/api/retail/p... Corrections: none as of Oct 18, 2026 Teach-only: setup and limits per Microsoft docs. Ticket to Tech is a made-up company; no real company data is used. On-screen cost: Azure's total is pending.