AI
From Zero to Working Agent in Minutes
Building a shopper assistant from scratch means picking a language, wiring a database, and configuring APIs. Todd Herman’s answer: skip it.
“it could actually take just 10 minutes to do this.” — Todd Herman
AI Foundry’s Templates page lists pre-built solutions with architecture diagrams, per-component pricing, and managed-identity security defaults. Clone the repo, run azd auth login, then azd up. Four minutes later the stack was live: an Azure OpenAI resource, an AI agent service, a container app, and log analytics. He walked through the shopper assistant live demo (09:13), querying camping gear and top-rated stoves against vectorized sample data loaded during deployment.
Swapping Data Without Touching Code
The template ships with camping product sample data. Varun’s company sells shoes.
“I don’t have to make any code changes.” — Todd Herman
Swapping data takes one step in the AI Foundry Agents page: open the deployed agent, find the knowledge base, upload your files, click Update. The platform vectorizes the documents and loads them into the search index automatically. Herman uploaded shoe inventory, then asked the agent for sneakers for a commute. Products came back immediately, grounded on the new data, with no infrastructure changes and no redeploy.
Authentication and Content Filtering
Sharing a container app URL with colleagues before adding access controls exposes the app to anyone with the link. Two settings close that gap.
Container Apps’ built-in Authentication tab lets you attach an identity provider in a few clicks. Herman selected Microsoft, set a 90-day client expiration, and Azure created the app registration and wired the identity connection automatically. For content safety, AI Foundry’s Guardrails page adds input and output filters with configurable thresholds, blocking jailbreak attempts without any code changes. Both settings applied in under two minutes, no redeploy required.
Multi-Agent Orchestration for Complex Tasks
A single-agent template handles focused retrieval. A marketing campaign needs multiple steps, multiple data sources, and human sign-off mid-run.
The multi-agent template deploys six agents on GPT-4o via Semantic Kernel. Herman walked through the multi-agent marketing campaign demo (20:44): a group chat manager routes work to a planner, which pauses to collect the audience (professionals aged 20-30) and a $50,000 budget. A product agent queries the database, a marketing agent writes social posts and email copy, a generic agent generates design files, and a procurement agent submits the print order.
Template Catalog and the Signal Behind It
Microsoft launched 10 AI Foundry templates at Build. The list came from analysis of 1,500+ customer conversations.
“we actually went through 1,500-plus customer conversations to figure out different signals based on where customers are most interested” — Todd Herman
Beyond the two demos, the catalog covers a content processing workflow (multimodal input: invoice photos, PDFs, text files) mapped to a structured schema via Azure AI Content Understanding, an enterprise deployment template following the Azure Well-Architected Framework with landing zones and encryption, and a document generation workflow that lets teams chat with their existing documents to draft contracts and statements of work. Each template is meant to be extended, not just run as-is.
Notable Quotes
it could actually take just 10 minutes to do this. Todd Herman · ▶ 1:42
I don’t have to make any code changes. Todd Herman · ▶ 13:39
we actually went through 1,500-plus customer conversations to figure out different signals based on where customers are most interested Todd Herman · ▶ 32:31
Key Takeaways
- An Azure AI Foundry template deploys a full agent stack — OpenAI, container app, search index — in about 4 minutes with azd up.
- Replacing sample data requires only a file upload in the AI Foundry agent editor; vectorization and indexing happen automatically.
- 10 templates launched at Build; the catalog was shaped by 1,500+ customer conversations identifying where teams get stuck.