What does it really take to ship an AI agent?

Microsoft Developer introduces “Sparkles”, a cupcake-ordering AI agent built with Azure AI Foundry and Microsoft Agent Framework, and follows the end-to-end path from building to operating the agent in production.

Overview

The video covers what it takes to ship an AI agent beyond a demo, using a concrete example agent (“Sparkles”) and walking through the full development and operations cycle.

The agent production problem

• Frames the gap between a working prototype and a production-ready agent.

Choosing from a wide array of models

• Connects the agent to a deployed model.
• Discusses selecting an appropriate model from multiple available options.

Building an agent in the portal

• Demonstrates building the agent directly in the Azure AI Foundry portal.

Building in code: Agent Framework and SDK

• Shows how to build the same agent in code using:
  • Microsoft Agent Framework
  • The associated SDK

Loading persona and UI assets from MCP prompts

• Loads the shop’s persona and a welcome banner from MCP prompts.

Calling live tools (Cupcake Store)

• Uses tool calling to invoke live “Cupcake Store” tools as part of the agent’s execution.

Inspecting execution traces

• Inspects execution traces to understand what the agent did and why.

Tracing tokens, cost, and latency

• Tracks operational signals that matter in production:
  • Token usage
  • Cost
  • Latency

Evaluating agent quality

• Evaluates the resulting conversations to assess agent quality.

Monitoring agents at scale

• Covers monitoring practices for running agents across broader usage.

Governance, security, and guardrails

• Discusses governance and guardrails needed to operate agents safely and consistently.

Video chapters

• 0:00 - The Agent Production Problem
• 0:36 - Choosing From a Wide Array of Models
• 0:56 - Building an Agent in the Portal
• 1:09 - Building in Code: Agent Framework & SDK
• 1:41 - Tracing Tokens, Cost, and Latency
• 2:29 - Evaluating Agent Quality
• 3:02 - Monitoring Agents at Scale
• 3:22 - Governance, Security, and Guardrails