microsoft/agentic-factory-hack

microsoft/agentic-factory-hack

Description: Build AI-powered predictive maintenance orchestration with Microsoft Foundry, multi-agent workflows, and MCP

Language: Python

License: MIT

Stars: 61

Forks: 381

Open issues: 22

Created: 2026-01-14T14:31:21Z

Pushed: 2026-06-16T04:22:24Z

Default branch: main

Fork: no

Archived: no

README:

Intelligent Predictive Maintenance Hackathon

Welcome to the Intelligent Predictive Maintenance Hackathon. In this workshop you’ll explore how AI agents can support predictive maintenance in a tire manufacturing setting—focused on practical patterns you can adapt to real environments.

Introduction

In this hackathon you’ll explore how agent workflows can support predictive maintenance: faster triage, more consistent diagnosis, and smoother handoffs from detection to repair planning.

Using a code-first approach with Microsoft Agent Framework, Microsoft Foundry, and Azure, you'll build a small multi-agent workflow that mirrors a real maintenance process: monitoring telemetry, identifying anomalies, performing fault diagnosis, planning repairs, and proposing a schedule. The goal is a clear, observable workflow that can be reasoned about, tested, and iterated on.

Learning Objectives

By participating in this hackathon, you will learn how to:

• Implement five focused agents (Anomaly Detection, Fault Diagnosis, Repair Planner, Scheduler, Orderer) and connect them to data/tools
• Use GitHub Copilot (including a custom agentplanning coding agent) to accelerate development while keeping changes reviewable
• Add persistent agent memory (threads) and apply it to maintenance scenarios where history matters
• Instrument the solution so you can inspect agent outputs, tool calls, logs, and traces during development
• Build an end-to-end agent workflow using Microsoft Agent Framework and sequential orchestration hosted in Aspire with agent-to-agent (A2A) communication.

Scenario

This hackathon uses a fictitious tire manufacturing company as an example, but the scenario is applicable to most manufacturing environments where machines are involved in production.

The image below illustrates the conceptual scenario. 

❶ The factory operates a production line made up of multiple specialized machines, each responsible for a distinct step in tire manufacturing (e.g., mixing, extrusion, curing, inspection).

❷ A key business metric is Overall Equipment Effectiveness (OEE)—how effectively equipment is utilized during planned production time. Unplanned downtime and quality losses directly reduce throughput and profitability.

❸ Maintenance technicians perform both *planned* (time/usage-based) maintenance and *condition-based* maintenance triggered by sensor readings and detected anomalies. Because machines are complex, tasks require specific skills, certifications, and safety procedures.

❹ Spare parts are held in a local warehouse with limited stock. When inventory falls below thresholds or parts are needed for a repair, replenishment is coordinated with external suppliers, often with lead times and ordering constraints.

❺ The production environment integrates with multiple core systems (e.g., WMS/ERP/CMMS/MES/QMS/knowledge bases). These systems coordinate production, maintenance execution, inventory, quality, and scheduling—making end-to-end orchestration essential.

The maintenance process is complex and requires coordination across people, parts, and systems. Common challenges include:

• Detecting when a machine needs attention, who should be notified, and how to prioritize the response.
• Diagnosing the root cause of irregular machine behavior using telemetry plus shared knowledge and historical patterns.
• Determining which parts are required, whether they are in local stock, and how to handle supplier lead times.
• Matching work to technicians with the right skills, certifications, and safety requirements.
• Scheduling technicians based on availability, shifts, and constraints (e.g., access windows, downtime limits).
• Finding the optimal maintenance window that minimizes production disruption while keeping risk acceptable.

Agents can help by coordinating data, decisions, and handoffs across these steps. The intent is not to “replace” technicians or planners, but to reduce time-to-triage and make the workflow more consistent and observable.

The image below shows the agent roles used in this hackathon and how they interact with telemetry and existing factory systems.


❶ Telemetry is emitted from machine sensors and evaluated for deviations (for example, simple thresholds or an anomaly detection model).

❷ The Anomaly Classification Agent classifies the situation (e.g., normal/warning/critical), enriches it with machine context, and raises an alert with an appropriate priority.

❸ The Fault Diagnosis Agent proposes likely root causes using telemetry plus documented historical knowledge (runbooks, prior work orders, technician notes).

❹ The Repair Planner Agent translates diagnosis into an actionable repair plan: required parts, estimated effort, and required skills/certifications. It can draft a work order for review.

❺ The Maintenance Scheduler Agent proposes a maintenance window based on priority/risk, technician availability, and production constraints.

❻ The Parts Ordering Agent checks inventory and suggests replenishment orders based on supplier lead times and constraints.

The factory process spans many systems, teams, and areas of responsibility. Agents closer to analytics might be built by one team, while agents tied to day-to-day operations might be built and operated by another. Teams also vary in skills, tooling, and technology stack preferences. This hackathon embraces that reality: we keep agents independent, but designed to collaborate through well-defined inputs/outputs and shared observability.

Architecture

In this hackathon you’ll build multiple agent applications (Python and .NET) and connect them using a sequential workflow in the Microsoft Agent Framework. The workflow itself is described in the scenario section above; the focus here is the supporting platform components.

The image below illustrates the architecture used throughout the hackathon.


❶ The...