Foundry Connected-Agent Orchestrator ⭐ OPTIONAL

Learning Objectives

• Use Azure AI Foundry's connected-agent pattern to orchestrate specialist agents without code.
• Register your Fabric Data Agents from Modules 7–8 as Foundry tools.
• Test cross-domain reasoning in the Foundry Playground with full trace visibility.

How It Works

Single-domain agents answer questions inside their own scope; multi-agent orchestration adds cross-domain reasoning. This is achieved without orchestration code - Azure AI Foundry provides a built-in connected-agents pattern in which a primary agent automatically delegates to specialist agents registered as tools.

The flow:

1. Open your pre-provisioned Foundry project at ai.azure.com.
2. Create a City Ops Orchestrator agent (gpt-4o-mini).
3. Attach your Fabric Data Agents from Modules 7–8 as connected agents.
4. Test in the Foundry Playground.

Build the Orchestrator

1. Open your Azure AI Foundry project

   Open a new browser tab → ai.azure.com. Sign in with your lab account.

   On the home screen pick your pre-provisioned Foundry project (each lab user has their own). See Module 0 · Setup & Environment Check for context.

   ⓘ

   Why Foundry? Azure AI Foundry is Microsoft's managed agent runtime. It provides a connected-agents pattern (one agent delegating to specialists as tools), built-in tracing, and a one-click publish path to Microsoft 365 Copilot, Teams, and the web - without orchestration code.

2. Create the City Ops Orchestrator agent

   In the left rail, click Agents → + New agent. Fill in:

   • Name: City Ops Orchestrator
   • Deployment: gpt-4o-mini (already in the project)
   • Instructions (paste exactly):

   Agent Instructions

   You are the City Ops Orchestrator - chief of staff for a Smart City operations team.
   
   For any user question, decide which of your connected specialist agents to call:
   - Hospital Ops Agent - patient vitals, room occupancy, movement
   - Transit Ops Agent  - train status, speed, delays
   
   You may call multiple specialists in the same turn when a question spans domains.
   After gathering specialist signals, synthesize ONE concise answer with:
   
     SITUATION   - what each specialist reported (one bullet per domain)
     ACTIONS     - recommended operational steps based on the specialist signals
     CONFIDENCE  - HIGH if all relevant specialists responded; MEDIUM otherwise
   
   Be terse and operational. Do not invent data the specialists did not return.

   Click Create.

3. Attach your Fabric Data Agents as connected agents

   On the agent canvas, scroll to Tools & Knowledge → + Add tool → Connected agent.

   For each Fabric Data Agent you built in Modules 7–8, add one connected-agent entry:

   Tool name	Description (what it does)	Endpoint
   hospital_ops	Answers questions about patient vitals, room occupancy, and movement on the hospital floor. Use for any clinical or capacity question.	Hospital agent endpoint (from your Fabric agent → Settings → Endpoint)
   transit_ops	Answers questions about train fleet status, speeds, and delays across Red/Blue/Green lines.	Transit agent endpoint (from your Fabric agent → Settings → Endpoint)

   ✓

   The description field is critical - the orchestrator's LLM uses it to decide when to call this tool. State precisely which questions each agent is qualified to answer.

   If you only built the Hospital agent in Module 8, attach that one alone - the orchestrator supports any number of connected agents.

4. Test in the Foundry Playground

   Click Try in Playground. Paste the canonical cross-domain question:

   ❓

   "Region A has train delays AND rising hospital load. What should the city do?"

   Inspect the trace pane on the right - the orchestrator calls hospital_ops and transit_ops in parallel, then synthesizes the response.

   Expected output shape:

   Playground Output

   SITUATION
   - Hospital: 4 critical patients, SpO2 trending down on PAT-10003.
   - Transit: Train-Green-1 delayed 8/10 readings; Green line 11 mph below baseline.
   
   ACTIONS
   1. Activate hospital staffing tier-2.
   2. Reroute Green-line passengers to the Blue line shuttle.
   3. Open incident bridge with transit + ED leadership.
   
   CONFIDENCE: HIGH (2/2 relevant specialists responded).

   Try a few additional prompts - the orchestrator selects which specialists to call automatically:

   • "How's hospital load looking right now?" → Hospital only
   • "What's the operational picture across the city?" → both agents
   • "Are there any concurrent emergencies?" → both agents

References

• Azure AI Foundry Agent Service - build and run connected-agent orchestrators.