Two renames and one rebuilt primitive
Lesson 1 of 5 in Microsoft Foundry Agent Service, Three Ways to Run an Agent.
Search for “Azure AI agents” and you will land in three different product eras at once — three names for the platform, two incompatible agent APIs, and a pile of blog posts that were accurate when written. Before any architecture, get the map straight.
Microsoft Learn documents the naming evolution explicitly: Azure AI Studio → Azure AI Foundry → Microsoft Foundry for the platform, and Azure Cognitive Services → Azure AI Services → Foundry Tools for the services portfolio. Current branding, and the one to use: Microsoft Foundry. The tell that nothing underneath was thrown away: the Azure resource type is still Microsoft.CognitiveServices/accounts.
The rename is cosmetic. The API change is not.
The original Foundry Agent Service copied the OpenAI Assistants shape: assistants, threads, messages, runs. The current service — the one Microsoft calls Agents v2 — is built on the Responses API and a different trio of primitives: agents, conversations, and responses. Old code does not limp along; it migrates or it stops.
| Classic (Agents v0.5 / v1) | Current (Agents v2) | What actually changed |
|---|---|---|
Assistants API | Responses API | The wire protocol. The Assistants API sunset on 26 August 2026, alongside retirement of the |
Threads / Messages / Runs / Assistants | Conversations / Items / Responses / Agent versions | A conversation is durable server-side history made of items — messages, tool calls, tool outputs. A response is the unit of execution. An agent is identified by name plus version, not a GUID. |
Hub + Azure OpenAI + Azure AI Services resources | One Foundry resource with projects | One resource, one project endpoint. Fewer moving parts to wire up, one place where governance and networking attach. |
Several SDK packages |
| You talk to the project endpoint with a standard OpenAI-shaped client for inference, and the projects SDK for platform concerns. |
Agents (classic) | Foundry Agents (GA) | Classic agents are deprecated with retirement on 31 March 2027. Microsoft documents no migration path from hub-based projects or Azure OpenAI Assistants — expect a rebuild, not a lift. |
Key terms: agent, agent loop, tool call, MCP, memory, orchestration
Two more orientation facts worth carrying into the next lesson.
Models are swappable. The catalog spans models from Microsoft, OpenAI, Anthropic, Meta and others — the docs said “more than 10,000” when this pack was checked, which you should read as an order of magnitude rather than a number — and Agent Service lets you change the model behind an agent without touching agent code.
GA is uneven, deliberately. The new portal is generally available, and so are core Agents, publishing to Microsoft 365 Copilot and Teams, Routines, and tracing for prompt and hosted agents. Workflows, the agent optimizer, tracing for workflow and external agents, and several networking experiences were still preview in September 2026. On this platform “preview” is a real engineering signal: no SLA, and not a place to put your revenue path.
Interactive checkpoint quiz (2 questions) — open this page in a browser to take it.