Microsoft Foundry: What's New in 2026 (formerly Azure AI Foundry)

If you've been building with "Azure AI Foundry" through 2024 and 2025, very little of your code has to change — but the docs, SDKs, examination questions (looking at you, AB-100), and Microsoft branding all shifted. This post is the practical update: what's the same, what's renamed, what's actually new, and what to learn first if you're catching up.

The rename: what it is and isn't

In January 2026 Microsoft rebranded Azure AI Foundry to Microsoft Foundry. The motivation, from Microsoft's product blogs: Foundry is no longer purely an Azure product. It's the platform that powers Copilot Studio agents, Microsoft 365 Copilot extensibility, Dynamics 365 AI features, and the Azure AI Apps and Agents Developer Associate certification's curriculum.

What didn't change:

• The underlying capabilities — model catalog, agent service, evaluations, prompt flows, content safety
• The Azure billing model — Foundry workloads still bill through Azure subscriptions
• The SDK class names — the Python and .NET SDKs continue to use AIProjectClient and related types
• Your existing deployments — no breaking changes, no migration required

What did change:

• The portal name and URL (now ai.microsoft.com alongside the existing ai.azure.com)
• Documentation namespace — learn.microsoft.com/azure/ai-foundry now redirects to learn.microsoft.com/foundry
• Certification content — AB-100's July 22, 2026 refresh uses "Microsoft Foundry" throughout
• Marketing positioning — Foundry is now pitched as the unified platform for the agentic Microsoft ecosystem

Practical impact: If your prep material or codebase still says "Azure AI Foundry", that's fine — it works. But for exam questions, blog posts, and team-wide vocabulary, use "Microsoft Foundry" going forward.

Foundry Agent Service: the new agentic core

The Foundry Agent Service is the most significant 2026 addition. It's a managed service for building, deploying, and operating agents — including multi-agent orchestrations — without writing the agentic-loop infrastructure yourself.

What it provides:

• Agent definition — declarative configuration of an agent's model, system prompt, tools, and behavior
• Tool integration — first-class support for Foundry Tools, custom OpenAPI tools, Code Interpreter, file search, and Model Context Protocol (MCP)
• Multi-agent orchestration — patterns for agents calling agents (Agent2Agent protocol support)
• Persistent threads and runs — managed state across user requests
• Observability — built-in telemetry, run inspection, and trace export to Application Insights
• Identity and access — integration with Microsoft Entra for authentication and authorization

If you've used the OpenAI Assistants API, the Foundry Agent Service is conceptually similar but tightly integrated with Microsoft's enterprise stack (Entra, Purview, Application Insights, Microsoft Sentinel). It's also model-agnostic — your agents can run on GPT-4o, Claude, Mistral, Phi, or any model in the catalog.

Model Catalog expansion

The Foundry Model Catalog has grown substantially through 2025 and 2026. As of mid-2026, key updates:

• Frontier models — GPT-4o family, GPT-4.1, Claude 4.x family (Opus, Sonnet, Haiku, Fable), Llama 3.x and 4.x, Mistral Large, Phi-4
• Specialized models — multimodal (vision + audio), code-specific, domain-specific (medical, legal, financial)
• Open-source models — hundreds of community models available with Foundry-managed inference
• Model router — a new capability that intelligently routes requests to the most suitable model based on the request characteristics. Critical for cost optimization at scale.

The model router deserves a callout — it's both a feature you can use directly and a concept the AB-100 exam tests. The router can balance latency, cost, and capability across a configured set of models, transparent to the calling code.

Evaluations framework

Foundry's evaluation tooling matured significantly in 2026. The framework now covers:

• Built-in evaluators — relevance, groundedness, coherence, fluency, similarity, retrieval accuracy
• Risk and safety evaluators — hate, sexual content, violence, self-harm, jailbreak resistance, indirect attack vulnerability
• Custom evaluators — your own code-based or model-based judges
• Evaluation runs — batch processing against test datasets with results stored for comparison
• CI/CD integration — evaluations as a gate in your deployment pipeline via the Foundry SDK

If you've struggled with "how do I know my AI app actually works?" — this is Microsoft's answer. Evaluation runs in Foundry are first-class artifacts, comparable across model and prompt changes.

Content Safety

Content Safety in Foundry now goes beyond simple content moderation. The 2026 capabilities:

• Prompt shields — defending against direct and indirect prompt injection (the latter is critical for agentic applications where models read tool outputs)
• Groundedness detection — flagging when a model's response isn't supported by provided context (RAG hallucination defense)
• Protected material detection — flagging when generated text reproduces copyrighted material verbatim
• Custom categories — define your own content classifications for industry-specific risks

For production agentic applications, prompt shields are especially important. Agents read tool outputs, web content, and user-provided data — all of which can contain adversarial instructions.

Prompt Flow and the AIProjectClient

Prompt Flow (Foundry's visual designer for LLM application pipelines) has expanded to handle agentic workflows. The Python and .NET SDKs continue to ship the AIProjectClient as the primary entry point — it now handles:

• Listing and instantiating agents
• Running and inspecting evaluations
• Querying the model catalog
• Managing connections (data sources, tools, MCP servers)
• Reading run telemetry

If you started building with Azure AI Foundry's earlier SDK in 2024, your code from then still works in mid-2026 — Microsoft has maintained backward compatibility through the rename and feature expansion. New code should use the agent-first APIs rather than the older fine-tuning-and-deployment-focused patterns.

What this means for your projects

If you're actively building on Foundry:

• Adopt the Agent Service for new agentic features — you save building the orchestration layer yourself
• Wire evaluations into CI — treat AI app quality the same way you treat any other code quality gate
• Enable prompt shields on any agent that consumes external content (web search, document loading, tool results)
• Consider the model router for cost optimization once you have multiple model use cases

If you're studying for AB-100:

• Use current Microsoft Foundry documentation, not older Azure AI Foundry material
• Know the Agent Service primitives — agents, runs, threads, tools, MCP integration
• Understand evaluations as an architectural concern, not just a quality concern
• Be ready for content-safety questions — they're up-weighted in the July 22, 2026 refresh

Related reading

• AB-100 exam guide — how Foundry topics map to the exam
• Is AB-100 worth it? — honest ROI breakdown for Microsoft architects
• The Microsoft AI cert ladder in 2026

The bottom line

Microsoft Foundry is more of an expansion than a reinvention. The Azure AI Foundry features you knew in 2025 are still there — but the platform now extends across Microsoft 365 and Dynamics, the Agent Service has matured into a credible production runtime, evaluations are a first-class artifact, and content safety covers the agentic-era threats. If you've been on the sidelines waiting for "the agent platform to mature", mid-2026 is when that wait ends.