Demo 5 — Multi-agent Workflow (Event Planning Workflow)

Reference:
- https://learn.microsoft.com/en-us/agent-framework/tutorials/workflows/agents-in-workflows?pivots=programming-language-python

This demo chains multiple specialized agents sequentially using a workflow (edge) to divide and integrate event planning tasks.

Objectives

• Split agents by role (specialization) and fix the order with Edges to build a stable pipeline
• Use Hosted tools (Web Search / Code Interpreter) and MCP tool (sequential-thinking) within the workflow
• (Optional) Output OTel spans in a single line to observe “which agent/tool was called”

The workflow in this demo follows this order:

1. coordinator (overall planning + sequential-thinking)
2. venue (venue research: Hosted Web Search)
3. catering (catering research: Hosted Web Search)
4. budget_analyst (cost estimation and allocation: Hosted Code Interpreter)
5. booking (final integration: outputs plan in markdown)

Prerequisites

Common (Foundry Agents)

• Demo 2 is complete (Foundry env vars are configured)
• agent-framework-foundry is installed
• az login completed (this demo uses AzureCliCredential by default)

Required env vars:

FOUNDRY_PROJECT_ENDPOINT="https://<account>.services.ai.azure.com/api/projects/<project-id>"
FOUNDRY_MODEL="<your-foundry-model-deployment-name>"

Additional (specific to this demo)

1) sequential-thinking (MCP stdio)

• node / npx must be available (MCPStdioTool launches npx ... locally)

2) Hosted Web Search (Bing connection)

This demo performs web searches for venues/catering, so a Bing grounding connection ID is required.

Set one of the following:

• BING_CONNECTION_ID (or BING_PROJECT_CONNECTION_ID)
• Or for Custom Search:
  • BING_CUSTOM_CONNECTION_ID + BING_CUSTOM_INSTANCE_NAME
  • (or BING_CUSTOM_SEARCH_PROJECT_CONNECTION_ID + BING_CUSTOM_SEARCH_INSTANCE_NAME)

src/demo5_workflow_edges.py explicitly loads the repository root .env and only fills in unset/empty env vars (mitigating the empty-string injection issue in Dev Container / Codespaces).

Steps

Step 1. Verify .env (recommended)

At minimum, the following are required:

• FOUNDRY_PROJECT_ENDPOINT
• FOUNDRY_MODEL
• BING_CONNECTION_ID (or an equivalent Bing connection setting)

Step 2. Log in (Entra ID)

az login --use-device-code

Step 3. Run

python3 -u src/demo5_workflow_edges.py

(Optional) To pause after execution:

DEMO_PAUSE=1 python3 -u src/demo5_workflow_edges.py

Expected behavior:

• Workflow Result: is displayed after Running workflow...
• The output appears as multiple steps (results from each specialist)
• (In environments with OTel enabled) Logs like [TOOL] ... tool=... / [AGENT] ... will be interspersed

Technical Notes (Design decisions)

• The workflow follows the pattern of “dividing event planning among multiple specialists, with booking integrating at the end”
• The pinned SDK in this repository uses FoundryChatClient(...).as_agent(...) instead of create_agent(...) for safety, so the implementation is based on as_agent

Troubleshooting

npx not found

• Install Node.js (usually not needed in the dev container)

No Bing connection

• Set BING_CONNECTION_ID (or BING_PROJECT_CONNECTION_ID)

403 / Forbidden

• Have you completed az login?
• Does the executing user have RBAC assigned on the Foundry Project / Hub?

Failed to resolve model info

• Is FOUNDRY_MODEL set to the deployment name from Models + endpoints in Foundry?

DNS resolution fails before starting

• Check your Private networking / private DNS configuration
  • The included script performs a DNS pre-check and provides a clear error message

Next Demo (DevUI)

In Demo 6, we visualize this “Event Planning Workflow” with DevUI. → Open demo6.md to continue.