Connected Agents & Multi-Agent

Multi-Agent Orchestration

Foundry supports two patterns for multi-agent systems:

1. Connected Agents (Hub-and-Spoke)

Register specialized agents as "tools" for an orchestrator agent. The orchestrator delegates tasks without custom routing code.

// Orchestrator agent with connected sub-agents:
orchestrator = project.agents.create_agent(
    model="gpt-5.4",
    name="Orchestrator",
    instructions="Route user requests to the appropriate specialist.",
    tools=[
        {"type": "connected_agent", "agent_id": search_agent.id},
        {"type": "connected_agent", "agent_id": analysis_agent.id},
        {"type": "connected_agent", "agent_id": writing_agent.id}
    ]
)

2. Multi-Agent Workflows

A stateful orchestration layer for complex, multi-step business processes. Maintains context and state across long-running tasks with approval gates and branching.

Microsoft Agent Framework v1.0 (April 2026)

On April 3, 2026, Microsoft released v1.0 of the Microsoft Agent Framework, officially merging AutoGen and Semantic Kernel into a single, unified open-source SDK for .NET and Python.

Agent Memory Service (Preview)

The Memory Service allows agents to persist context across multiple sessions without custom databases:

• User Profile Memory — Stores user preferences (dietary restrictions, language, etc.) across interactions
• Chat Summary Memory — Distilled summaries of topics covered in past conversations
• Scoped Access — Memory is segmented per-user for secure, isolated experiences
• Free in Preview — No additional cost during preview; you pay only for underlying model usage