Design stateful agentic loops with Microsoft Foundry Agent Service

Why it matters

Where an agent’s state lives decides what can resume, branch and be audited, and the Agents v2 model in Foundry Agent Service changes that answer from v1’s threads and runs.

What to practise

Define a versioned prompt agent with PromptAgentDefinition and azure-ai-projects 2.x, and keep multi-turn state in a conversation. Run a reflection loop whose iteration limit belongs to the application, then fork an alternate path with previous_response_id outside the conversation. Compare the supplied v1 thread-and-run code with the v2 replacements. In the Foundry portal, check the agent’s version history and match each response ID to its trace.

How it fits

The conversation and the response branch are two of 1.2.5’s persistence choices; the retained IDs without prompt text are 1.2.7’s execution evidence. The application-owned iteration limit is the bounded loop from 1.1.5.