Azure AI Foundry Before Ignite 2025
Ahead of Microsoft Ignite 2025, we take a sober look at Azure AI Foundry: what the Agent Service delivers as a hosted runtime since its GA, which roadmap signals sit behind the agent factory narrative, why the Azure prefix is eroding, and which capabilities engineering teams should adopt now or continue to watch.
The platform question before Ignite
Microsoft Ignite 2025 takes place from 18 to 21 November at the Moscone Center in San Francisco, with a pre-day on 17 November. For teams building agent systems on Azure, that date marks a decision point. The platform is changing quickly: Azure AI Studio became Azure AI Foundry at Ignite 2024, the Agent Service reached general availability at Build in May 2025, and the Agent Factory blog series has been preparing the next wave since August. The practical question is which dependencies are already defensible today.
Platform churn carries a real cost. Names change, SDKs merge, and preview services are reshaped or retired. Committing early delivers capability sooner but may create migration work later. This article records the platform as of 20 September 2025 and separates what we consider safe to adopt from what we would only watch. Our criterion is not novelty but reversibility: how expensive would it be to leave a component if Ignite changes it?
What Azure AI Foundry is today
Azure AI Foundry is Microsoft's platform for building AI applications and agents. It was introduced on 19 November 2024 at Ignite as the successor to Azure AI Studio. Today it combines a model catalog of more than 11,000 models with an SDK, the Foundry Agent Service, and observability tooling. The catalog spans Microsoft-hosted and partner models under one deployment and billing model. OpenAI's GPT-5 has been available there since 7 August 2025, alongside a model router that selects a model variant per request.
Two boundaries define the product. Foundry is not an orchestration framework: Semantic Kernel and AutoGen sit below it, and Microsoft announced at Build 2025 that both are merging into a single SDK. It is also not a no-code tool; that role belongs to Copilot Studio. Foundry targets professional developers who want managed infrastructure underneath code they still control.
The agent factory narrative
Microsoft's framing for all of this is the agent factory. The phrase appeared as a session title at Build 2025 — "Azure AI Foundry: The Agent Factory" — and since 13 August 2025 it names a six-part series on the Azure blog: use cases and design patterns, a first agent, observability, the path from prototype to production, open standards, and security.
A factory promises standardized parts and repeatable assembly. The phrase is marketing; the underlying concerns are not. Tracing, evaluation, interoperability, and governance determine whether agents survive production. We therefore read the series mainly as a roadmap signal: Microsoft explicitly names the themes it intends to make concrete at Ignite.
The series also fixes a vocabulary: knowledge tools versus action tools, observability as a first-class pillar, the open agentic web as the target picture. Shared vocabulary is not a small thing. It is what lets a platform vendor, its partners, and its customers argue about the same objects.
Hosted runtimes as managed state
The Foundry Agent Service, generally available since 19 May 2025, is a hosted runtime. Threads, runs, tool invocations, and file state live on the server side. Built-in tools cover Azure AI Search, Code Interpreter, File Search, function calling, Bing grounding, and OpenAPI; MCP connects external tool servers, and the A2A protocol exposes agents to other agents.
A hosted runtime reduces infrastructure work, but it neither designs prompts nor defines quality. Observability records traces; you still have to establish the evaluation criteria. Lock-in also remains: thread state and tool configuration live inside the service, while responsibility for keeping that state exportable stays with you.
The alternative remains self-hosted orchestration: Semantic Kernel or AutoGen processes on your own compute, with state in your own store. That costs operational work and buys portability. For regulated workloads whose conversation data must stay inside defined boundaries, this trade-off deserves an explicit decision rather than a default.
Naming drift as a signal
One detail from Build 2025 deserves separate attention: the naming. The local runtime shipped as Foundry Local — no Azure prefix — in preview on Windows and macOS, built on ONNX Runtime. Windows Copilot Runtime became Windows AI Foundry. Documentation increasingly says Foundry Models and Foundry Agent Service, and Microsoft's developer blog for the platform lives under devblogs.microsoft.com/foundry.
Microsoft renames platforms at Ignite. Azure AI Studio became Azure AI Foundry there in November 2024, one year after its own launch. A brand that now spans cloud, Windows, and on-device runtimes has outgrown its cloud prefix. We expect the Azure AI part of the name to disappear; the open question is when.
Adopt now or watch
Our rule is simple: adopt what is generally available and rests on open protocols; watch what is in preview or bound to a single vendor surface. Applied to the Foundry stack as of 20 September 2025, this yields the following picture.
Two judgments deserve explanation. Although parts of the MCP tooling remain in preview, we adopt MCP behind our own abstraction because the open protocol has traction beyond Microsoft and will survive any platform rename. We only watch A2A orchestration. The multi-agent patterns are months old, and Ignite is likely to reshape both APIs and pricing before the service stabilizes.
| Capability | Status on 20 Sep 2025 | Our call |
|---|---|---|
| Foundry Agent Service | GA since 19 May 2025 | Adopt for stateful single agents |
| OpenAPI and function tools | GA | Adopt |
| MCP tool connections | Preview | Adopt behind your own abstraction |
| A2A multi-agent orchestration | Preview | Watch until after Ignite |
| Foundry Local | Preview on Windows and macOS | Watch and evaluate for edge scenarios |
| Unified SK and AutoGen SDK | Announced at Build 2025 | Watch |
Outlook toward November
What do we expect on 18 November? A consolidated brand, plausibly plain Foundry without the Azure prefix. Hosted multi-agent workflows as the headline. Deeper agent identity and governance building on the Microsoft Entra Agent ID introduced at Build. Tighter coupling between Copilot Studio and the Foundry runtime. None of this is confirmed; all of it follows the pattern of the last twelve months.
Some predictions will be wrong in detail; our stance at Blue IT Systems does not depend on them. Build against protocols rather than product names, keep orchestration behind your own interfaces, preserve exportable conversation state, and treat every preview feature as replaceable. A November rename should trigger a configuration change in your system, not a migration project. The broader direction — agents as managed cloud workloads — remains unaffected.
Sources
- Microsoft Build 2025: The age of AI agents and building the open agentic web (Microsoft Official Blog, 19 May 2025)
- Announcing General Availability of Azure AI Foundry Agent Service (Microsoft Tech Community, 19 May 2025)
- Agent Factory: The new era of agentic AI — common use cases and design patterns (Microsoft Azure Blog, 13 August 2025)
- Unlock Instant On-Device AI with Foundry Local (Microsoft Foundry Blog, 19 May 2025)
- Microsoft Opens Registrations for Ignite 2025 in San Francisco (Thurrott, 25 June 2025)