Microsoft Foundry

The managed AI platform on Azure. Foundry handles the infrastructure you don't want to build: model deployment, agent hosting, knowledge indexing, safety filtering, and observability. You write the agent logic; Foundry runs, scales, and governs it.

Rebranded from Azure AI Studio in early 2026. One Foundry resource replaces the old Hub + Azure OpenAI + AI Services setup.

Resource Model

One Foundry resource governs everything. Projects isolate teams - separate deployments, separate agents, same billing and compliance boundary.

---

Foundry Models

1,900+ models behind one Azure endpoint. Prompts and completions are never used to train models. Data residency is configurable per deployment.

PTU sizing rule: size to P80 traffic, overflow to PAYG. PTU is cheaper per token at sustained load but you pay whether you use it or not.

# Production pattern - Managed Identity, no API keys
from azure.identity import DefaultAzureCredential
from openai import AzureOpenAI

credential = DefaultAzureCredential()
token = credential.get_token("https://cognitiveservices.azure.com/.default")

client = AzureOpenAI(
    azure_endpoint="https://my-foundry.openai.azure.com/",
    azure_ad_token=token.token,
    api_version="2024-12-01-preview"
)

response = client.chat.completions.create(
    model="gpt-5",
    messages=[{"role": "user", "content": "Explain RAG in 3 sentences."}]
)

---

Foundry Agent Service

Managed runtime for agents. Three types:

Memory (preview) - three types that persist across runs:

• Procedural - agent learns how to do the work across runs. Early benchmarks show +7-14% absolute success rate gains
• User - retains user preferences and facts across sessions
• Session - maintains thread context within a conversation

Toolboxes (preview) - one managed MCP endpoint for all tools. Configure once, expose to any MCP client. Connects to Foundry IQ, Work IQ, Fabric IQ without custom plumbing.

Routines (preview) - schedule agents on a timer. Overnight triage, daily reporting, periodic compliance checks.

Voice Live (GA) - real-time voice for agents. STT, TTS, turn detection, interruption handling, avatars in one API.

from azure.ai.projects import AIProjectClient
from azure.identity import DefaultAzureCredential

client = AIProjectClient(
    credential=DefaultAzureCredential(),
    endpoint="https://my-foundry.services.ai.azure.com"
)

# Prompt agent - declarative
agent = client.agents.create(
    model="gpt-5",
    name="support-agent",
    instructions="You are a tier-1 support agent. Use the knowledge base to answer questions. Escalate billing issues to the billing agent.",
    tools=[
        {"type": "file_search"},
        {"type": "connected_agent", "connected_agent": {"id": billing_agent.id}}
    ]
)

# Publish to Teams
# Done via Foundry portal or Copilot Studio publish pipeline

Publish targets: Teams, Microsoft 365 Copilot, Copilot Studio, web apps - via a single governed publish pipeline.

---

Foundry IQ

The knowledge layer. Foundry IQ replaces custom RAG pipelines - ingest from multiple sources, retrieve with SLA-backed latency, ground agents without custom indexing glue.

Data pipeline features: layout-aware chunking (understands document structure, not just text splits), image extraction and serving, permission sync via Purview (ACLs from source propagate to retrieval), sensitivity label inheritance.

OneLake as knowledge source: Connect Foundry IQ directly to OneLake. Index files stored there, expose as knowledge inside agents.

---

Foundry Tools

Pre-built AI APIs that extend agents. All available as Toolbox tools.

Gotcha: Foundry Tools bill separately from model inference. Budget for both when building document processing or voice agents.

---

Observability

Automatic when Application Insights is connected. Zero code changes for Foundry-hosted agents.

Traces export to Application Insights (default), Datadog, or any OTLP backend.

Evaluation tools - all target different stages of the dev cycle:

---

Security & Guardrails

Important: Content Safety is not optional in production. Prompt Shield alone catches most jailbreak attempts. ACS adds deterministic rule enforcement for high-stakes scenarios.

---

Key Takeaways

• Use Managed Identity in production - API keys are a rotation and audit liability.
• PTU is cheaper per token at sustained load, but you pay whether you use it or not. Size to P80 and overflow to PAYG.
• Foundry IQ replaces custom RAG pipelines - permission sync via Purview and layout-aware chunking are the two things hardest to rebuild yourself.
• Content Safety and Prompt Shield are not optional in production. Enable both from day one.
• Hosted Agents (GA July 2026) are the production hosting target for MAF-built agents - sandboxed, stateful, framework-agnostic.
• Toolboxes (preview) consolidate all tool connectivity into one MCP endpoint - reduces integration surface significantly in multi-tool agents.