Foundry Agent Service

Managed agent runtime for building production AI agents. Define agents declaratively or run your own code in Foundry's sandboxed environment.

Agent Types

Prompt Agents

Define what the agent does, what tools it has, and what knowledge it accesses. Foundry handles orchestration, state, and execution.

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

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

agent = client.agents.create(
    model="gpt-5",
    name="support-agent",
    instructions="You are a customer support agent.",
    tools=[
        {"type": "file_search"},
        {"type": "code_interpreter"}
    ]
)

Quickstart: Create a prompt agent

Hosted Agents

Your Python/JS code runs in Foundry's sandboxed runtime with dedicated compute, memory, filesystem access, and state. Framework-agnostic - supports Microsoft Agent Framework, LangGraph, CrewAI, or custom.

Two protocols:

• Responses API - OpenAI-compatible stateful interactions
• Invocations protocol - schema-free pass-through

from azure.ai.agents import AgentsClient
from azure.identity import DefaultAzureCredential

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

Quickstart: Deploy a hosted agent

Connected Agents (Multi-Agent)

Agents delegate to other agents based on intent.

billing_agent = client.agents.create(
    model="gpt-5", name="billing-agent",
    instructions="Handle billing inquiries.",
    tools=[{"type": "function", "function": {"name": "get_invoice"}}]
)

triage_agent = client.agents.create(
    model="gpt-5", name="triage-agent",
    instructions="Route requests to specialists.",
    tools=[{"type": "connected_agent", "connected_agent": {"id": billing_agent.id}}]
)

Connected agents guide

Memory

Agents retain context across interactions. Three types (public preview):

Memory documentation

Routines

Run agents on a schedule (public preview). No user trigger required.

Use cases: overnight issue triage, daily reporting, periodic data sync, scheduled compliance checks.

Voice Live

Real-time voice for agents - speech recognition, text-to-speech, turn detection, interruption handling, and avatars in one API.

Voice Live documentation

Toolboxes

Single managed endpoint for all agent tools (public preview). Configure tools once, point any MCP client at one URL.

• Tools, skills, MCP clients under one governed endpoint
• Tool search helps select relevant tools per task
• Connects to Foundry IQ (Work IQ, Fabric IQ, Web IQ)

Toolboxes documentation

Publishing

Deploy agents where users already work:

Identity, permissions, and policy flow through automatically.

Copilot Studio Integration

Copilot Studio integration guide