What Is Microsoft Foundry? Azure's Agent Platform Explained
What Is Microsoft Foundry? (May 2026)
Microsoft Foundry is Azure’s unified platform for building, deploying, and operating AI agents and applications — portal, runtime, identity, memory, evals, multi-model catalog, all in one place. Originally Azure AI Foundry in November 2024, rebranded to Microsoft Foundry in early 2026 as Microsoft positioned it as the enterprise agent operating system. Here’s what it actually does today.
Last verified: May 25, 2026.
TL;DR
- What it is: Azure’s unified multi-model agent platform.
- Origin: Launched Nov 2024 as Azure AI Foundry, rebranded to Microsoft Foundry early 2026.
- Model catalog: OpenAI, Anthropic, Mistral, Microsoft (MAI), open models.
- Runtime: Foundry Hosted Agents (refreshed April 22, 2026) with hypervisor isolation, Entra agent identity, scale-to-zero pricing.
- Orchestration: Microsoft Agent Framework + Copilot Studio + Foundry multi-agent workflows.
- Identity: Entra agent identity (Microsoft’s identity standard for agents).
- Evals: M365 Copilot Evaluations Tool + (since May 20, 2026) RAMPART/Clarity open-source companions.
- Best for: Azure-native, M365-integrated, enterprise compliance-focused agent deployments.
What Microsoft Foundry actually does
Foundry is best understood as a layered platform with five capabilities:
1. Multi-model catalog
You get a single API surface across:
- OpenAI — GPT-5.5 family, including reasoning + instant tiers.
- Anthropic — Claude Opus 4.7 and Sonnet 4.6 (Azure-hosted).
- Mistral — Large, Codestral, and other Mistral SKUs.
- Microsoft (MAI) — Microsoft’s own model family.
- Open models — Llama 5, Qwen 3.6, DeepSeek V4 (curated subset).
The value: model choice without model-vendor lock-in. You can route by task without writing separate integration code for each vendor.
2. Foundry Hosted Agents (runtime)
The April 22, 2026 refresh added the production primitives enterprises needed:
- Hypervisor isolation — each agent runs in a hypervisor-isolated environment, addressing tenant-isolation concerns that blocked earlier Foundry adoption.
- Entra agent identity — agents get first-class Entra identities, enabling consistent IAM across humans + agents.
- Scale-to-zero pricing — pay for active agent runtime, not idle capacity.
- Built-in memory — agent memory primitives integrated with the runtime.
- Tool gateway — secure tool execution with identity scoping.
3. Orchestration (three flavors)
| Flavor | Use case | Audience |
|---|---|---|
| Microsoft Agent Framework | Production agents in code (C#, Python) | Engineers |
| Copilot Studio | Low-code agents in M365 surface | Analysts, citizen devs |
| Foundry portal multi-agent | Drag-and-drop orchestration | Architects, ops |
All three deploy onto the same Foundry Hosted Agents runtime and share identity, evals, and observability.
4. Evals and safety
- M365 Copilot Evaluations Tool — quality evals for M365-bound Copilot agents.
- RAMPART (open-source, May 20, 2026) — agent safety CI test framework.
- Clarity (open-source, May 20, 2026) — pre-build assumption audit.
- Built-in observability — Foundry portal traces, metrics, agent-level dashboards.
5. Multi-agent workflows
Foundry’s drag-and-drop multi-agent designer lets you compose lead-agent + sub-agent patterns visually, using any model from the catalog for each role. The runtime handles inter-agent state, retries, and observability.
What you can actually build with Foundry
Three patterns dominate in May 2026:
Pattern 1: Production line-of-business Foundry agents
Engineers build agents in Microsoft Agent Framework (C# or Python), deploy to Foundry Hosted Agents, integrate with Azure data sources and Microsoft 365, monitor via Foundry observability. Pattern matches LangGraph + custom runtime in the open-source world, but with Microsoft’s identity, governance, and ops baked in.
Pattern 2: Copilot Studio agents inside M365
Analysts and citizen developers use Copilot Studio’s low-code experience to build agents that live inside Teams, Outlook, Word, Excel. The agent has access to M365 data through Microsoft Graph, executes inside the M365 surface, and is governed by M365 Copilot Evaluations + Microsoft Purview. Pattern matches “internal automation app builder” but with deep M365 integration.
Pattern 3: Multi-agent workflows in Foundry portal
Architects design lead-agent + sub-agent compositions in Foundry’s drag-and-drop designer. The lead agent dispatches sub-agents (potentially using different models from the catalog for each role) and orchestrates results. Useful for complex cross-domain workflows that don’t fit cleanly inside a single agent or inside the M365 surface.
Where Foundry fits in the broader enterprise AI stack
| Layer | Example tools |
|---|---|
| End-user agents | M365 Copilot (built on Foundry), ChatGPT Workspace |
| Vertical agents | Salesforce Agentforce, ServiceNow AI Agents |
| Process orchestration | Camunda ProcessOS, Power Automate |
| Infrastructure + runtime | Microsoft Foundry, AWS AgentCore, Claude Managed Agents |
| Models | GPT-5.5, Claude Opus 4.7, Gemini 3.5 Flash, etc. |
Foundry is the infrastructure + runtime layer for Azure. Enterprises typically pair it with vertical agent platforms (Agentforce, ServiceNow) for industry-specific workflows.
The May 2026 production-readiness picture
Three recent moves complete Foundry’s enterprise story:
| Move | Date | What it added |
|---|---|---|
| Foundry Hosted Agents refresh | Apr 22, 2026 | Hypervisor isolation, Entra agent identity, scale-to-zero |
| Microsoft Agent Framework 1.0 | April 2026 | Production SDK GA in C# and Python |
| RAMPART + Clarity open-sourced | May 20, 2026 | Agent safety CI testing + assumption audit |
Together they close the gaps that were blocking serious enterprise production adoption through 2025. As of May 2026, “Foundry is not production-ready” is no longer a defensible objection inside Azure-anchored organizations.
Who Microsoft Foundry is (and isn’t) for
Strong fit:
- Azure is your primary cloud.
- M365 Copilot is core to your productivity story.
- You want one platform across multiple models with consistent governance.
- Entra-based identity is your standard.
- You value opinionated end-to-end developer experience over framework freedom.
Weak fit:
- AWS or GCP is your primary cloud.
- You’re committed to LangGraph, CrewAI, or another framework and don’t want to migrate.
- True multi-cloud parity is a hard requirement.
- Open-source-first philosophy is non-negotiable.
Honest caveats
- Multi-cloud story is thin. Foundry is strongly Azure-anchored. The cross-cloud bridges exist but the native experience is Azure.
- Microsoft Agent Framework is opinionated. Teams already invested in LangGraph or CrewAI may find migration painful.
- Multi-agent orchestration is real but younger than Claude Managed Agents’ shipping equivalent (Anthropic’s multi-agent orchestration shipped at Code with Claude on May 6, 2026).
- Pricing is multi-dimensional. Model tokens, runtime hours, hosted agents, memory, tools — the bill has several axes. M365 Copilot license bundles cover in-scope agents but net-new workloads can surprise on cost.
How Foundry connects to OpenAI, Anthropic, and Microsoft’s model strategy
A key insight for 2026: Microsoft’s AI strategy is not “OpenAI exclusively.” Foundry deliberately surfaces Anthropic and Mistral as first-class options alongside OpenAI and Microsoft’s own models. The bet is that enterprises want model choice without integration tax — Foundry provides the choice + the unified integration.
This positions Microsoft well for the multi-model agent era and reduces dependency on any single model provider’s roadmap.
Verdict
- What it is: Azure’s unified, opinionated agent platform — multi-model, multi-orchestration, deeply integrated with Entra and M365.
- Who needs it: Azure-native, M365-deep enterprises serious about production agents.
- What it competes with: AWS Bedrock AgentCore (composable AWS alternative), Claude Managed Agents (opinionated Anthropic-only), Google Vertex AI Agents (Google equivalent).
- What it’s ready for in May 2026: Production deployments — the April 22 refresh and May 20 RAMPART/Clarity launch closed the prior gaps.
- Bottom line: If you’re Azure-native, Microsoft Foundry is now the most production-ready unified agent platform available. Outside Azure, the calculus changes.