Agents as a New Programming Control Structure

▶ Watch (05:41)

Before agentic AI, every application flow was hard-coded. Rules, while loops, fixed buttons. Now a model, a set of instructions, and a collection of tools form a new control structure. That structure converts natural language into control flow. Give an agent “plan an event” and it decides whether to find a venue, send calendar invites, or generate social media, depending on context. The agent does not know the sequence in advance.

That shift moves decision-making from if-statements into prompts, which makes reliability the central engineering problem for anyone building these systems.

Live Voice-Controlled Multi-Agent Demo

▶ Watch (11:40)

Amanda asked the Voice interface to plan Seth’s birthday for 10 people in Seattle on a $100 budget. The BuildEvents live birthday party demo (11:40) ran end to end on stage: Bing Custom Search found a venue, a Logic App sent calendar invites, an image agent generated a custom graphic, and the full post went live on LinkedIn in front of the audience.

The scenario had changed up to a week and a half before the keynote.

“we made an agentic app that doesn’t care.” — Seth Juarez

Swapping to a new use case required only one updated prompt for the Voice controller.

Building Agents with Azure AI Foundry Agent Service

▶ Watch (24:19)

Azure AI Foundry Agent Service ships in three setup modes. Basic wraps the OpenAI Assistants API and adds Azure-specific tools: Bing Grounding, SharePoint, Azure AI Search. Standard brings your own storage so all processed data stays in resources you can see in the Azure portal. The third mode adds a private virtual network for enterprises that need it.

“Azure AI Foundry runs on trust, not trends.” — Amanda Foster

All three tools attach in two to three lines of code each. Amanda walked through the connection steps in the single agent LinkedIn post live demo (29:32), ending with a live post to LinkedIn.

Evaluation Data Drove the Move to Specialized Agents

▶ Watch (33:03)

The team started with one generalist agent handling venue search, content writing, and LinkedIn posting. Elijah ran a GPT-4o-generated test set through four metrics in the eval comparison: generalist vs specialized agents (34:22): intent resolution, tool call accuracy, task adherence, and response completeness. The generalist was faster but scored lower on every task-specific benchmark. A dedicated Social Media Content Writer agent scored significantly higher at the cost of more tokens.

One agent with one clear instruction outperforms one agent with fifty. Splitting the generalist into task-focused sub-agents gave the Voice controller cleaner targets and produced consistent outputs.

Orchestrating Sub-Agents via Implicit Function Calling

▶ Watch (41:20)

The Voice controller is itself an agent. Seth converted each Foundry sub-agent into a tool call with two parameters: additional_instructions and query. A continuation on the same function call ID acts as a back channel, letting the controller report “still working” mid-run. Each sub-agent holds thread-local state; the controller thread sees only inputs and outputs. Seth called this implicit agentic orchestration. The full code is in the Voice controller function call architecture walkthrough (46:20).

OpenTelemetry tracing colored Voice inputs yellow and agent calls blue. When a Voice activity bug appeared, one session trace gave engineers everything they needed to fix it.

Four Lessons for Production Agentic Apps

▶ Watch (54:14)

Four questions before shipping. First: is the prompt right? Amanda showed a spreadsheet of Voice controller prompt iterations, revised until the agent reliably called the correct sub-agent. One changed instruction sometimes fixed the entire system. Second: are the tools right? Third: did you pick the right agentic pattern? The BuildEvents top-level controller delegating to task-specific sub-agents was a choice driven by Voice modality. Fourth: will your organization actually let you put this in production?

Seth open-sourced the full codebase live on stage.

Notable Quotes

Azure AI Foundry runs on trust, not trends. Amanda Foster · ▶ 18:23

we made an agentic app that doesn’t care. Seth Juarez · ▶ 11:30

we called it the Phat agent, P-H-A-T, Seth Juarez · ▶ 36:20

Key Takeaways

  • Specialized agents with narrow prompts and consistent inputs/outputs score significantly higher on evals than one generalist agent.
  • Prompt refinement is the first debug tool. A single changed instruction can fix broken orchestration across the whole system.
  • Azure AI Foundry’s three setup modes let teams start in minutes and add full enterprise data sovereignty when the org requires it.