Episode Details
Back to Episodes
Azure AI Foundry Multi‑Agent Systems: Planning, Collaboration, Tooling That Don’t Nuke Prod
Season 1
Published 4 months, 2 weeks ago
Description
(00:00:00) The Power of Multi-Agent Systems
(00:00:32) The Limitations of Single-Agent Systems
(00:02:32) Introducing Multi-Agent Systems
(00:03:55) Roles and Responsibilities in Multi-Agent Systems
(00:04:47) Building with Azure AI Foundry and Semantic Kernel
(00:09:50) Demo Scenario 1: Device Cleanup in Intune
(00:13:38) Demo Scenario 2: Zero-Touch Onboarding
(00:17:17) Demo Scenario 3: Automated Security Hardening
(00:22:58) Best Practices for Multi-Agent Systems
(00:25:06) Closing Thoughts and Call to Action
In this episode of M365.fm, Mirko Peters builds a real multi‑agent system with Azure AI Foundry and Semantic Kernel that can plan, execute, and verify changes across Intune, Entra ID, and Microsoft Graph — without turning your tenant into a lab experiment.
WHAT YOU WILL LEARN
Most “AI agent” demos collapse the entire help desk, change board, and postmortem into one over‑prompted bot — and then act surprised when context, cost, and safety fall apart.
Multi‑agent systems fix this by splitting work into roles: one agent plans, one executes with tools, one reviews changes, and one talks to humans.
Instead of a single giant prompt, you get small, deterministic loops where each agent sees only what it needs and every risky action goes through tools with RBAC and logging.
This episode argues that real enterprise AI is not about a smarter chatbot — it is about building a digital team that behaves like a disciplined operations crew.
WHY MULTI‑AGENT SYSTEMS WITH AZURE FOUNDRY WORK
(00:00:32) The Limitations of Single-Agent Systems
(00:02:32) Introducing Multi-Agent Systems
(00:03:55) Roles and Responsibilities in Multi-Agent Systems
(00:04:47) Building with Azure AI Foundry and Semantic Kernel
(00:09:50) Demo Scenario 1: Device Cleanup in Intune
(00:13:38) Demo Scenario 2: Zero-Touch Onboarding
(00:17:17) Demo Scenario 3: Automated Security Hardening
(00:22:58) Best Practices for Multi-Agent Systems
(00:25:06) Closing Thoughts and Call to Action
In this episode of M365.fm, Mirko Peters builds a real multi‑agent system with Azure AI Foundry and Semantic Kernel that can plan, execute, and verify changes across Intune, Entra ID, and Microsoft Graph — without turning your tenant into a lab experiment.
WHAT YOU WILL LEARN
- Why a single “do‑everything” agent breaks down in real enterprise environments
- How to design Planner, Operator, Reviewer, and Messenger agents with clear roles and boundaries
- How to wire agents into real tools: Intune, Entra ID, Graph API, Azure Automation, and Log Analytics
- How a multi‑agent workflow can cut time‑to‑fix from 12 minutes to 3 minutes on real incidents
- How to treat tools as “hands” and memory as a budget, not a magic black box
- How to use Azure AI Foundry to define agents, tools, knowledge, and safety policies
- How to keep RBAC, PIM, logging, and Zero Trust intact while agents do the work
Most “AI agent” demos collapse the entire help desk, change board, and postmortem into one over‑prompted bot — and then act surprised when context, cost, and safety fall apart.
Multi‑agent systems fix this by splitting work into roles: one agent plans, one executes with tools, one reviews changes, and one talks to humans.
Instead of a single giant prompt, you get small, deterministic loops where each agent sees only what it needs and every risky action goes through tools with RBAC and logging.
This episode argues that real enterprise AI is not about a smarter chatbot — it is about building a digital team that behaves like a disciplined operations crew.
WHY MULTI‑AGENT SYSTEMS WITH AZURE FOUNDRY WORK
- Planner focuses on intent and constraints; Operator focuses on tools and execution; Reviewer focuses on safety and compliance; Messenger handles approvals and communication
- Tools are explicit: Graph, Intune, Automation runbooks, Functions, Logic Apps, and RAG via Azure AI Search
- Azure AI Foundry manages threads, safety, evaluations, and tool wiring so you don’t hand‑roll orchestration
- Semantic Kernel gives you planners, skills, function catalogs, retries, and cancellation baked into code
- Managed Identities, split RBAC, and PIM keep permissions tight and auditable
- Log Analytics, Application Insights, and content safety give you full traceability of every tool call
- One giant agent is a gas‑station Swiss Army knife: looks capable, bends on the first serious job
- Multi‑agent design = roles, boundaries, and parallelism mapped to real operational responsibilities
- Keep prompts short and move real power into well‑designed tools with strict schemas
- Treat memory as a constrained resource