A multi-agent route optimization system using Microsoft Agent Framework with Group Chat pattern. Automatically redistributes delivery routes when a driver becomes unavailable.

                    ┌──────────────┐
                    │ Orchestrator │ ◄── Azure OpenAI (GPT-4o)
                    └──────┬───────┘
               ┌──────────┼──────────┐
               ▼          ▼          ▼
          ┌───────┐  ┌────────┐  ┌────────┐
          │ Route │  │Schedule│  │ Driver │
          │Analyzer│ │Optimizer│ │Assigner│
          └───────┘  └────────┘  └────────┘

Agents: RouteAnalyzer (identifies affected routes) → ScheduleOptimizer (plans redistribution) → DriverAssigner (matches drivers) → Orchestrator (synthesizes plan)

• Azure Developer CLI (azd)
• Python 3.11+
• Azure subscription with Azure OpenAI access

azd auth login
azd env new dev
azd env set AZURE_LOCATION eastus2
azd up

A web-based UI is included at src/frontend/index.html for interacting with the optimization system.

Features:

• Visual agent workflow progress
• Real-time chat interface showing agent conversations
• Driver selection dropdown
• Markdown-rendered agent responses

Running locally: Open src/frontend/index.html in a browser after configuring config.local.js with your API endpoint.

After deployment: The postprovision hook generates config.local.js with your Azure Function URL.

cd src/functions
python -m venv .venv && .venv\Scripts\activate
pip install -r requirements.txt
# Update local.settings.json with AZURE_OPENAI_ENDPOINT
func start

├── azure.yaml              # azd configuration
├── infra/                  # Bicep IaC (OpenAI, Functions, Monitoring)
├── src/
│   ├── functions/          # Python Azure Functions (API + Agents)
│   └── frontend/           # Web UI (HTML + Tailwind CSS)
└── hooks/                  # azd lifecycle hooks

Azure OpenAI (GPT-4o) • AI Foundry Hub & Project • Azure Functions • Application Insights • Storage Account

Azure OpenAI usage is logged to Log Analytics. See Azure OpenAI Monitoring Queries for KQL queries to track:

• Token usage - Prompt and completion tokens over time
• API calls - Total calls per model, streaming vs non-streaming
• Latency - Per request and per model (avg, max, P95)
• Request details - Individual request inspection, slow request detection

azd down

MIT