The first architectural decision is not the model. It's the agent type!

Foundry Agent Service now offers three types. Each one determines how much you build and how much the platform manages for you.

1. Prompt agents: the fastest path (easy to build, but limited)

You define instructions, a model and tools, and Foundry runs the agent. There's no application code, no containers and no infrastructure. Author it in the portal, or define it through the SDK/REST API so it fits into CI/CD.

Best for: internal tools and production agents that don't need custom orchestration.

2. Voice-based prompt agents: real-time conversation (Speech → Model → Phone)

Same configuration-first approach, but for spoken interaction. You connect over a real-time WebSocket through Voice Live, which handles speech recognition, turn-taking and speech synthesis. You can choose native speech-to-speech for low latency, or a cascaded STT → LLM → TTS pipeline for more control. Telephony via Teams Phone or Twilio.

Best for: customer-facing voice and call-center scenarios. Note: still in preview.

3. Hosted agents: full control (most powerful, no limits)

Bring your own code with Agent Framework, LangGraph, the OpenAI Agents SDK, the Anthropic Agent SDK, the GitHub Copilot SDK or plain custom code. Ship it as a container or a .zip, and Foundry runs it with a managed endpoint, autoscaling, a dedicated Entra identity, session state and end-to-end observability.

Best for: agents that call into your own code, custom orchestration and multi-agent systems.

My rule of thumb

Start with a prompt agent. If you hit the limit (and you will), daenet can help you with a hosted agent.

I will host the AI Lab ILL335-CGN – Building & deploying AI agents with Microsoft Foundry at the Microsoft AI Tour in Cologne on October 22, 2026.

MicrosoftFoundry #AIAgents #AgenticAI #Azure #MCP #AgentFramework

Originally published on LinkedIn.