MCP & Semantic Kernel AI Agents: Building IT Ops Automation That Actually Takes Action

(00:00:00) Transforming AI from Talker to Worker
(00:00:40) The Shift from Q&A to Action
(00:01:50) The Three Ingredients of AI Orchestration
(00:04:30) The Six Parts of a Capable IT OPS Agent
(00:10:08) Microsoft Stack: The Containment Field
(00:16:45) Blueprint I: SK Planner + Graph via MCP
(00:22:32) Blueprint II: Azure Open AI Tool Calling with Managed Identity
(00:27:40) Blueprint III: Incident Autoremediation and IT OPS
(00:35:28) The Power of Guardrails and Responsibility
(00:41:48) Key Takeaways and Future Directions

In this episode of M365.fm, Mirko Peters shows how to turn AI from chatty assistant into a disciplined IT Operations agent that plans, executes, verifies, and stays inside governance—from Semantic Kernel and MCP to Azure OpenAI with Managed Identity.

WHAT YOU WILL LEARN

• Why “chatbots that give advice” are wasting your AI potential compared to agents that actually act
• How to move from Q&A loops to a closed‑loop cycle: Intention → Plan → Tool Use → Result → Self‑Check → Next Step
• How real SRE teams wire agents to handle incidents end‑to‑end before a human even wakes up
• How MCP exposes tools like Microsoft Graph, Intune, App Insights, and internal APIs in a standard, discoverable way
• How Semantic Kernel turns those tools into orchestrated plans with sequential, parallel, and graph‑shaped tasks
• How Azure OpenAI with Managed Identity keeps credentials out of prompts and enforces RBAC at the tool boundary
• How to design stable “agent molecules” with persona, memory, planner, tools, policy, and verifier working together

THE CORE INSIGHT

Most AI projects stall at “better answers.” The real value appears when agents are allowed to do work in a closed loop with tools, checks, and guardrails.
MCP makes your tools visible, Semantic Kernel orchestrates the plan, Azure OpenAI reasons about steps, and Managed Identity constrains what’s actually allowed.
Instead of magic prompts, you get small, testable workflows where every action is logged, validated, and reversible.
This episode argues that serious AI in Microsoft shops is not about smarter chat—it is about building verifiable, identity‑bound agents that behave like cautious SREs.

WHY MCP + SEMANTIC KERNEL + MANAGED IDENTITY WORK

• MCP standardizes tool exposure so Graph, Intune, Service Health, and internal services describe themselves via schemas—not ad‑hoc plugins
• Semantic Kernel wraps MCP tools as functions, builds JSON payloads, and handles planning across multiple steps and branches
• Azure OpenAI uses tool‑calling while Managed Identity decides what each tool is actually allowed to do
• High‑risk actions (rollback, redeploy, bulk changes) require explicit approvals encoded in tools, not “pretty please” in prompts
• Audit envelopes and telemetry turn every tool call into evidence you can review, replay, or red‑team

KEY TAKEAWAYS

• Agents need six parts to stay predictable: persona, memory, planner, tools, policy, and verifier
• Start with narrow, high‑value flows like post‑deployment incident handling or password reset automation
• Put power in tools and identity scopes, not in giant prompts and hidden capabilities
• Measure success in MTTR reduction, ticket deflection, burnout reduction, and audit quality—not just “AI usage”
• Treat safety as physics: split Managed Identities, hard schemas, approval tokens, immutable logs, and scope‑drift monitoring

WHO THIS EPISODE IS FOR

This episode is ideal for SREs, platform engineers, IT operations teams, and cloud architects who want AI to fix real incidents, not just summarize them.
If you’re under pressure to “use AI” but worried about production safety, this conversation gives you a blueprint for governed, observable, and identity‑bound IT Ops agents on the Microsoft stack.

TOPICS COVERED

• From chatbots to acting agents in IT Operations
• MCP as the standard wiring for tools across Graph, Intune, App Insights, and internal APIs
• Semantic Kernel planning patterns and the six‑part agent molecule
• Azure OpenAI tool‑calling with Managed Identity for safe execution
• Blueprints for auto‑remediation, password reset, and post‑deploy incident handling
• Guardrails: approvals, identity splits, logging, red‑teaming, and model rotation strategies

ABOUT THE HOST

Mirko Peters is a Microsoft 365 consultant and digital workplace architect focused on building safe, observable automation on the Microsoft cloud.
Through M365.fm, Mirko shares blueprints, governance patterns, and real‑world stories that help IT and SRE teams turn AI agents into reliable colleagues instead of new risk surfaces

Become a supporter of this podcast: https://www.spreaker.com/podcast/m365-fm-modern-work-security-and-productivity-with-microsoft-365--6704921/support.