Skip to content

FAQ.md

Latest commit

 

History

History
243 lines (138 loc) · 7.37 KB

FAQ.md

File metadata and controls

243 lines (138 loc) · 7.37 KB

Frequently asked questions

General

What is this sample for?

It teaches patterns for building AI agents with:

  • Microsoft Agent Framework for agent logic
  • MCP for tool extensibility
  • .NET Aspire for multi-service orchestration
  • Multiple LLM provider support

See learning objectives.

Who is this for?

  • .NET developers building AI agents
  • Architects designing multi-service AI apps
  • Anyone exploring Agent Framework or Aspire
  • Engineers looking at deployment patterns for agent systems

Can I use this in production?

The architecture patterns are solid for production. A few things to think about first:

  • Review security settings (content filters, authentication)
  • SQLite won't scale forever — plan for Azure SQL or Cosmos DB under load
  • Add proper error handling and monitoring for your use case

How is this different from other chatbot samples?

It goes further than a demo:

  • Extensible via MCP (not hard-coded tools)
  • Works with multiple LLM providers
  • Explains the reasoning behind design choices
  • Actually deploys to Azure with azd up

Microsoft Agent Framework

What is Microsoft Agent Framework?

A .NET library for building AI agents. Gives you structured instructions, tool calling, multi-agent orchestration, and OpenAI-compatible APIs.

Official docs

How is this different from Semantic Kernel or AutoGen?

Feature Agent Framework Semantic Kernel AutoGen
Language .NET, Python .NET, Python, Java Python
Focus Production agents AI orchestration Multi-agent research
Hosting Web APIs Embedded Standalone
AG-UI Protocol Yes No No

Agent Framework is optimized for deployable web services. Semantic Kernel is more of a general orchestration library. AutoGen focuses on multi-agent research.

Can I use multiple agents?

Yes. The framework supports handoff orchestration (sequential chain with specialists), agent-as-tools (coordinator calls helpers), and single-agent mode.

Switch between them with the AgentMode setting in apphost.settings.json.

See multi-agent guide.

What's the AG-UI protocol?

A standard for agent-to-UI communication. It means you can swap frontends or agent implementations without rewiring everything.

Learn more

Model Context Protocol (MCP)

What is MCP and why use it?

MCP is a protocol for connecting AI agents to external tools and data sources. Tools become reusable across agents and frameworks, language-agnostic (Python tools in .NET agents), and independently deployable.

When should I use MCP vs. inline tools?

Use MCP when the tool is complex, reusable, or owned by a different team. Use inline tools for trivial, agent-specific functions or when you need to avoid the network hop.

Can I use existing MCP servers?

Yes. The MCP Server Registry has database connectors, API integrations (Slack, GitHub, Jira), file system access, and more.

How do I build my own MCP server?

See Tutorial 2. Short version: create a .NET project, add ModelContextProtocol.Server, implement tools with [McpServerTool] attributes, and map the /mcp endpoint.

LLM providers

Why is Foundry the recommended provider?

It bundles model routing (automatic model selection for cost/quality), content safety, monitoring, and evaluation into one platform. It also has a 99.9% uptime SLA.

See Foundry setup.

Can I use Azure OpenAI instead?

Yes, just change the config:

"LlmProvider": "AzureOpenAI"

See Azure OpenAI setup.

Can I use GitHub Models?

For development, yes:

"LlmProvider": "GitHubModels"

Rate limits make it unsuitable for production.

See GitHub Models setup.

Can I use OpenAI Platform (not Azure)?

Not currently supported, but adding it is straightforward:

  1. Extend LlmResourceFactory.cs
  2. Add OpenAI client configuration
  3. Update configuration schema

Can I use local models (Ollama, LM Studio)?

Not directly. You could point to an OpenAI-compatible endpoint by modifying LlmResourceFactory, but local models often have weak tool-calling support.

Aspire

What is Aspire?

A .NET framework for building cloud-native apps. Handles service orchestration, discovery, observability (logs, traces, metrics), and deployment to Azure.

Official docs

Do I need Aspire to use Agent Framework?

No. Aspire makes multi-service dev easier (service discovery, observability), but you can run the agent standalone. You'd just need to manage MCP connections, service URLs, and monitoring yourself.

Can I deploy without Aspire/azd?

Yes, deploy containers directly:

docker compose up  # Local development

# Or deploy to Azure manually
az containerapp create ...

But you'll lose Aspire's benefits (service discovery, observability).

Deployment

How do I deploy to Azure?

azd up

That's it. See the README.

Can I deploy to AWS or GCP?

Not with azd, but the app is just containers. You could deploy to ECS/Fargate or Cloud Run — you'd configure networking and environment variables manually.

What about scaling?

Container Apps auto-scales on HTTP request count (including scale-to-zero). For storage, swap SQLite for Azure SQL. MCP servers scale independently.

Customization

How do I change the interview flow?

Edit the agent instructions in AgentDelegateFactory.cs. See Tutorial 3.

Can I use this for other domains?

Yes. Change the agent instructions, swap in domain-specific MCP servers, and adjust the UI. The same patterns work for support bots, sales assistants, tutors, or intake forms.

How do I add new capabilities?

Preferred: create a new MCP server with tools, register it in AppHost, and connect it from the agent. For trivial agent-specific functions, define them inline.

See Tutorial 2.

Can I change the UI framework?

Yes. The agent exposes standard APIs, so you can replace Blazor with React, build a mobile app, or use Electron. The agent doesn't care about the frontend.

Troubleshooting

Where do I get help?

  1. Search Issues
  2. Open New Issue
  3. Stack Overflow (tag: microsoft-agent-framework)

Contributing

How can I contribute?

See CONTRIBUTING.md.

Can I use this in my project?

Yes, MIT licensed. Use it, modify it, ship it. Attribution is appreciated but not required.

Learning resources

Where should I start?

  1. Run the sample
  2. Learning objectives
  3. Architecture overview
  4. Tutorials

Still stuck? Open an issue.