AI-500: Designing and Implementing Multi-Agent AI Solutions

Note: This lab pre-provisions Azure resources at start — allow up to 10 minutes for the environment to become ready before beginning the exercises.

Turn the Ridgevault Financial multi-agent architecture from Lesson 1's whiteboard into a working code baseline. You'll sketch the five-specialist mesh (Portfolio Analyst, Compliance Officer, Risk Assessor, Client Relations, Investment Researcher) as a Mermaid diagram in docs/architecture.md, scaffold the Microsoft Agent Framework project against a Foundry-hosted gpt-5 deployment, drop empty agent stubs for the four specialists that later lessons will fill in, and wire the Portfolio Analyst end-to-end with a stub tool so it can answer a real Ridgevault portfolio question. Ridge — Ridgevault's in-app architect — pairs with you in the sidebar.

Note: This lab pre-provisions Azure resources at start — allow up to 10 minutes for the environment to become ready before beginning the exercises.

Ridgevault Financial runs a multi-agent advisory platform. In this hands-on lab you fix two problems on the investment_researcher agent: bare-model generic answers, and unaudit-able "industry source" citations. You will (1) write an advanced persona + few-shot system prompt that enforces Ridgevault's House View voice, (2) build a dynamic-context builder that trims the running thread to a token budget while preserving system + first/last user messages, (3) index Ridgevault's seed research library into Azure AI Search, (4) wire Foundry IQ RAG grounding into investment_researcher, and (5) run an end-to-end Ridgevault portfolio question that requires document grounding to answer correctly.

Note: This lab pre-provisions Azure resources at start — allow up to 12 minutes for the Foundry account, project, gpt-5 deployment, Storage account, Application Insights, and Function App (Flex Consumption) to become ready before beginning the exercises.

Ridgevault Financial's advisor stack depends on two tools nearly every advisor turn needs: a market snapshot for the assets in a portfolio, and a compliance check that flags positions above policy limits. In this hands-on lab you build both as a single fastmcp server, wire it into an Azure Function so any agent in any framework can call it over HTTP, and then attach the deployed MCP endpoint to Ridgevault's investment_researcher and compliance_officer agents via HostedMcpTool. By the end you will have shipped the exact tool-server pattern the AI-500 exam expects — portable, discoverable, HTTP-hosted MCP tools that multiple agents share.

Note: This lab pre-provisions Azure resources at start — allow up to 12 minutes for the environment to become ready before beginning the exercises.

Take Ridgevault Financial's four specialist agents — portfolio analyst, risk assessor, compliance officer, brief writer — and wire them into a nightly portfolio-review workflow two ways. First you build a fast sequential pipeline with Microsoft Agent Framework's SequentialBuilder on a shared conversation thread. Then you rebuild the same flow as a LangGraph StateGraph with typed per-node state and a human-in-the-loop approval gate at the compliance node — the graph pauses on interrupt() and resumes only when a human decides approve or reject. You capstone with graceful degradation: fault-inject one specialist and add an on-graph error edge that keeps the workflow completing.

Note: This lab pre-provisions Azure resources at start — allow up to 12 minutes for the environment to become ready before beginning the exercises.

Ridgevault Financial's Portfolio Analyst agent is drafting client-facing quarterly reviews under a regulated-communications banner. Compliance cannot sign off on general availability without three things: automated quality scores on outputs, a repeatable synthetic test bed that grows over time, and a human-in-the-loop review round for edge cases. In this hands-on lab you run Foundry's built-in evaluators (Relevance, Groundedness, Coherence) over a seed test set, add a custom LLM-as-judge for Ridgevault's regulatory-tone rubric, generate synthetic cases via the Foundry Simulator, push results to the Foundry portal and trigger a human review round, then wire the whole thing into a CI-style regression gate.

Note: This lab pre-provisions Azure resources at start — allow up to 12 minutes for the environment to become ready before beginning the exercises.

Instrument Ridgevault Financial's sequential portfolio-review multi-agent flow (Portfolio Analyst -> Risk Assessor -> Compliance Officer) with OpenTelemetry. Ship traces to Application Insights and Foundry Tracing, attribute prompt+completion tokens per agent turn, apply two cost optimizations to the noisy risk_assessor (prompt compression + LRU response caching), and set a drift alarm on the compliance_officer output distribution as a proxy for behavior drift. Grounded in the Ridgevault Financial scenario and guided by Ridge, the AI-500 platform reliability engineer.

Note: This lab pre-provisions Azure resources at start — allow up to 15 minutes for the environment to become ready before beginning the exercises.

Ridgevault Financial's multi-agent portfolio-review platform runs across a Foundry account and a fleet of downstream Container Apps. You are the platform's Zero Trust engineer for the day. In this lab you apply the Zero Trust model to the stack: attach a system-assigned managed identity to the Foundry account and grant it only the data-plane role it needs, move a Ridgevault internal API secret out of code and into Azure Key Vault (referenced from an env var via Key Vault reference syntax), and then run an adversarial safety scan against the portfolio-review flow to uncover prompt-injection and PII-leakage weaknesses. You close by writing two mitigations — a prompt-injection defense middleware and a PII redaction filter — and re-running the scan to confirm the risk score dropped.

Note: This lab pre-provisions Azure resources at start — allow up to 12 minutes for the Foundry account + project + gpt-5 deployment + Content Safety + Language accounts to become ready before beginning the exercises.

Wire a production-grade four-layer guardrail stack around Ridgevault Financial's portfolio-review flow, then gate every future change on a CI-style red-team regression. You will start from an unprotected Ridgevault Advisor agent and add, in order: (1) an Azure AI Content Safety + Prompt Shields input filter on every user message, (2) a tool-argument middleware that validates every portfolio_lookup account-number against the Ridgevault ^RV-[0-9]{8}$ regex plus a 7-year lookback rule, (3) a tool-result middleware that uses Azure AI Language PII entity recognition to redact SSN, DOB, phone, address, and person names before the model ever sees the tool return, (4) a response-layer middleware that runs Foundry Groundedness detection plus a deterministic regulatory-phrase blocklist (guaranteed return, risk-free, guaranteed to grow, insider information) on the model's drafted answer. Finally you run a labeled red-team JSONL set (direct-injection, tool-arg injection, PII exfiltration, regulatory, hallucination, and benign-control categories) through the assembled flow and inspect the per-category pass rate against Ridge's compliance-officer promotion thresholds. Everything is keyless (DefaultAzureCredential + Foundry User + Cognitive Services User + Cognitive Services OpenAI User).