Stop wiring infrastructure, start building agents.
Microsoft Developer uses Sparkles, a cupcake-ordering agent, to show how to choose between prompt and hosted agents in Microsoft Foundry. The walkthrough also covers tool configuration, reusable skills, and handling larger orders.
Start with a prompt agent
The example rebuilds Sparkles as a prompt agent with three core components:
- A model
- Instructions
- Cupcake Store MCP tools
These are defined once and versioned in Microsoft Foundry. This approach avoids managing containers, servers, or a deployment pipeline for the agent, rather than starting every project by wiring up scaling, state, and authentication infrastructure.
Configure tools and reusable skills
The walkthrough introduces the Copilot Harness, configures its tools, and adds reusable agent skills. The cupcake-store example expands from complex orders to bulk and party orders, including processing orders from CSV data.
It also covers configuring the agent with YAML and defining skills and MCP tools.
Video chapters
- 0:00 — Scale complex cupcake orders
- 0:34 — Introduce the Copilot Harness
- 1:07 — Configure harness tools
- 1:38 — Add reusable agent skills
- 2:44 — Enable bulk and party orders
- 3:18 — Process orders from CSV data
- 4:25 — Configure the agent with YAML
- 4:59 — Define skills and MCP tools
- 5:32 — Scale without managing infrastructure
Switch to a hosted agent when needed
A prompt agent is the starting point when a model, instructions, and tools are enough. When the implementation requires custom orchestration or a different framework, the same agent can move to a hosted agent.
Callers continue to invoke the agent by name, keeping the invocation consistent even as the implementation changes.