Architecting Multi-Agent Solutions with Copilot Studio and M365 Agents SDK

Session Date: May 20, 2025
Duration: 1 hour
Venue: Build 2025 Conference - BRK176
Speakers: Salem Bacha (Principal Architect, Microsoft), Gary Pretty (Principal Product Manager, Copilot Studio, Microsoft)
Additional Speakers: Vid Chari, Matthew Barbour, Mike Stall, Sarah Critchley
Link: [Microsoft Build 2025 Session BRK176]

Multi-Agent Orchestration

Executive Summary

This comprehensive session demonstrates Microsoft’s unified multi-agent architecture vision, showcasing seamless integration between Copilot Studio’s low-code agent development and the M365 Agents SDK’s pro-code capabilities. The team reveals how organizations can build sophisticated agent ecosystems that span platforms, languages, and development approaches, with live demonstrations of cross-platform agent orchestration, Visual Studio Code integration, and enterprise-grade multi-agent workflows that transform business processes.

Key Topics Covered

?? 1. The Multi-Agent Vision: Copilot vs Agents

Foundational Distinction

Vid Chari’s Strategic Framework: > “What we believe is that every employee will have a copilot… We also believe that every business process will be transformed by agents. So, Copilot is aligned to every employee, and agents are aligned to business processes.”

Core Differentiation:

  • Copilot - Personal assistant grounded in individual data (emails, chats, meetings, documents)
  • Agents - Business process automation aligned to organizational workflows
  • M365 Copilot - UI to AI, the interface for all agent interactions
  • Agent ecosystem - Programs using AI to automate and execute business processes

The Billion Agent Future

IDC Projection and Market Reality:

  • 1.3 billion agents by 2028 - IDC industry analyst prediction
  • Organizational diversity - Agents built by process owners, subject matter experts, and enterprise teams
  • Top-down and bottom-up - Both strategic initiatives and grassroots process improvement
  • Microsoft platform advantage - Direct integration with UI to AI across all Microsoft tools

??? 2. Microsoft’s Comprehensive Agent Platform

The Platform Advantage

Unified Development Ecosystem:

Microsoft Agent Development Stack:
??? Low-Code: Copilot Studio (visual canvas, rapid authoring)
??? Pro-Code: Visual Studio + GitHub (custom control, technical expertise)
??? Foundation: Azure AI Foundry (models, tools, orchestration)
??? Integration: M365 Copilot (universal UI to AI)

Strategic Benefits:

  • Enterprise-grade security - Governance, compliance, and data protection
  • Latest model access - Cutting-edge AI capabilities and orchestration
  • Company data integration - Seamless connection to Microsoft platform data
  • Cross-platform orchestration - Agents working together regardless of development approach

Development Approach Selection

Choice Framework:

  • Process owners/SMEs - Use Copilot Studio for visual, rapid development
  • Professional developers - Use pro-code tools for custom control and enterprise features
  • Hybrid teams - Combine approaches based on organizational needs and expertise
  • Common underpinnings - Knowledge, tools, and models shared across all approaches

?? 3. Copilot Studio Multi-Agent Capabilities

Traditional vs Multi-Agent Architecture

Current Agent Anatomy:

Traditional Agent:
??? Orchestrator (Generative, GA worldwide)
??? Knowledge sources
??? Tools and actions
??? Topics (high control workflows)

Enhanced Multi-Agent Architecture:

Multi-Agent System:
??? Lightweight specialized agents (focused, portable)
??? Connected external agents (cross-platform integration)
??? Orchestrator coordination (automatic agent selection)
??? Shared context and conversation history

Lightweight Specialized Agents

Gary’s Implementation Philosophy: > “We’re introducing the ability for you to build smaller, lightweight, focused, specialized agents inside of your agent itself… these move with your agent.”

Key Features:

  • Portable integration - Move between environments with parent agent
  • Instruction-driven - Natural language behavior definition instead of manual node creation
  • Tool sharing - Reference parent agent tools with granular control
  • Question capabilities - Built-in tools for asking questions and sending messages

?? 4. Live Demonstration: Contoso Banking Agent

Multi-Agent Banking Scenario

Contoso Checking Agent Architecture:

Banking Agent System:
??? Balance Information Agent (lightweight specialized)
??? Lost/Stolen Card Agent (complex conditional logic)
??? Mortgage Agent (connected Copilot Studio agent)
??? Car Loan Agent (connected Azure AI Foundry agent)

Balance Information Agent

Simple Instruction-Driven Workflow:

Instructions:
1. Fetch the user's profile and get list of user accounts
2. If user didn't specify account, ask which one to check
3. Use Get balance information tool to fetch and display data

Live Demo Results:

  • Natural language interaction - “What is the balance of my account?”
  • Dynamic questioning - Agent asks for account specification when needed
  • Follow-up support - “What about my savings account?” seamlessly handled
  • Tool restriction - Shared tools with granular usage control

Lost/Stolen Card Agent

Complex Conditional Processing:

Advanced Instructions:
1. Get list of user accounts and ask user to choose
2. Format account list with specific presentation rules
3. Freeze the account and get transaction list
4. Show last five transactions and check for suspicious activity
5. If suspicious found, call dispute endpoint for flagged transactions

Sophisticated Workflow Results:

  • Multi-step coordination - Chaining multiple tools between user interactions
  • Conditional logic - “I didn’t make the ATM withdrawals” triggers dispute process
  • Natural language processing - Conversational responses throughout complex workflow
  • Activity summary - Complete execution tracking and transparency

?? 5. Cross-Platform Agent Integration

Copilot Studio to Copilot Studio Connection

Salem’s Agent-to-Agent Demo:

  • Mortgage Agent - Specialized agent built with lightweight architecture
  • Connection settings - Simple toggle for agent availability
  • Publishing requirement - Agents must be published to be connectable
  • Conversation history control - Full context vs. private mode options

Implementation Simplicity: > “You as agent builders, you don’t have to worry about how do these agents talk to each other? How do these agents communicate? We’ve actually handling all this behind the scenes for you.”

Azure AI Foundry Integration

Cross-Platform Orchestration:

Foundry Agent Connection:
??? Agent Name: Car Loan Agent
??? Description: "Use this agent to answer car loan issues"
??? URL: Foundry hosting endpoint
??? Agent ID: Unique identifier from Foundry
??? Connection String: Authentication and access configuration

Authentication Options:

  • User authentication - End-user credentials passed through
  • Copilot author authentication - Service account for simplified demo scenarios
  • Multi-turn support - Full conversation continuity across platforms

?? 6. Multilingual Capabilities and Localization

Global Agent Deployment

Gary’s Multilingual Announcement: > “This week, we have made public preview available for additional language support for the generative orchestrator. You can now enable generative orchestration for any language supported by Copilot Studio.”

Revolutionary Localization:

  • 29 languages available - Complete language support in public preview
  • Automatic translation - Agent instructions in English, conversations in any language
  • Seamless experience - No manual translation required for agent logic
  • GA timeline - General availability within weeks

Live Spanish Demonstration

Technical Achievement:

  • English instructions - Agent built entirely in English
  • Spanish conversation - Full conversation flow automatically translated
  • Lost card scenario - Complex workflow operating flawlessly in Spanish
  • Zero additional work - Click language boxes and publish globally

?? 7. Visual Studio Code Extension for Copilot Studio

Pro-Developer Integration

Mike Stall’s Platform Extension: > “We now have a Visual Studio Code extension for Copilot Studio. It’s just gone into preview. You can go to the marketplace and download this today.”

Professional Development Features:

VS Code Integration:
??? Clone/Push workflow - Git-like operations with cloud sync
??? IntelliSense support - Full semantic understanding of agent structure
??? Error detection - Meaningful error messages and corrections
??? Offline development - Full local development capability
??? File-based editing - YAML files with rich IDE support
??? Source control integration - Change tracking and collaboration

Development Workflow

Seamless Cloud Integration:

  1. Clone agent - Download cloud agent to local YAML files
  2. Local development - Full IntelliSense, go-to-definition, error detection
  3. Offline capability - Work without internet connectivity
  4. Push changes - Sync modifications back to cloud
  5. Live validation - Changes appear immediately in Copilot Studio

Advanced Capabilities:

  • Semantic understanding - VS Code understands agent structure and relationships
  • Color-coded changes - Visual diff highlighting additions and modifications
  • Built-in topic management - Enable/disable system topics programmatically
  • Multi-language solution - Solve localization challenges through code

??? 8. M365 Agents SDK: Pro-Code Integration

SDK Philosophy and Architecture

Sarah Critchley’s Development Vision: > “The Agents SDK is an open source SDK from Microsoft, available in C#, JavaScript, and Python, and it allows you to build agents where you get to choose your AI model or services, your orchestrator, and your knowledge for grounding.”

Complete Developer Control:

SDK Architecture:
??? AI Models: Choose any service or provider
??? Orchestrator: Custom logic and decision-making
??? Knowledge: Custom grounding and data sources
??? Conversation Management: Built-in state, storage, auth
??? Channel Deployment: 15+ channels out of the box

Multi-Channel Deployment

Universal Agent Distribution:

  • Microsoft Teams - Native integration and deployment
  • Slack - Cross-platform business communication
  • M365 Copilot - Direct integration with Microsoft’s AI interface
  • 15+ channels - Comprehensive platform support
  • Custom channels - Build native application integrations

Organizational Integration Patterns

Enterprise Development Reality: > “It’s very common that our customers… will work in organizations where there’ll be teams of people… Some will be building agents in Copilot Studio. Others will be building agents in Visual Studio, in C#, in Python.”

Integration Strategies:

  • Dispatcher/broker patterns - Main SDK agent coordinating Copilot Studio specialists
  • Leverage existing work - Use Copilot Studio agents from custom applications
  • Multi-agent architectures - Complex orchestration across development approaches
  • No rebuild required - Implement SDK client in existing applications

?? 9. Technical Implementation: SDK Client Demo

Copilot Studio Client Integration

Matthew Barbour’s Technical Deep Dive:

// SDK Client Configuration
var client = new CopilotStudioClient();
await client.StartAsync();
await client.StartConversationAsync();
await client.AskAsync("Hi, who are you?");

Implementation Simplicity:

  • Minimal dependencies - Only core communication layer required
  • Connection string - Simple configuration from Copilot Studio
  • Event handling - Full eventing infrastructure access
  • Multi-platform - C#, JavaScript, Python support

Advanced Integration Patterns

Enterprise-Grade Features:

  • Semantic Kernel integration - Embedded client for advanced orchestration
  • HTTP client control - Custom networking and connection management
  • Authentication management - Flexible auth configuration options
  • Activity injection - Send adaptive cards and complex events

Development Options:

  • Direct connection URL - Simple single-agent integration
  • Environment ID - Multi-agent switching capabilities
  • Schema name - Dynamic agent discovery and connection
  • Multi-cloud support - Different environments and regions

??? 10. Advanced Features: Triggers and Autonomous Operations

Event-Driven Agent Activation

Salem’s Trigger System: > “You can also have the agent listen to specific triggers… You have multiple types of trigger that you can now have your specialized agent listen for and get activated on.”

Trigger Types:

  • Conversation start - Agent activation on session initiation
  • Orchestrator decision - AI-driven agent selection
  • External events - System-triggered autonomous operations
  • Custom triggers - Business-specific activation conditions

Autonomous vs Conversational Operations

Gary’s Implementation Scope: > “Every single thing that you’ve seen today works just as well for autonomous scenarios as well. Multi-agent works across them both.”

Operational Modes:

  • Conversational - User utterance triggered agent workflows
  • Autonomous - External event or system triggered operations
  • Hybrid coordination - Mixed human and system initiated processes
  • Multi-modal integration - Various activation and coordination patterns

Session Highlights

“Multi-agent systems don’t need to be hard.” - Vid Chari

“We want to give you all the tools in your toolbox to scale out your solutions in a way that makes sense for you.” - Gary Pretty

“You as agent builders, you don’t have to worry about how do these agents talk to each other? We’ve actually handling all this behind the scenes for you.” - Salem Bacha

“We want to help you… make use of agents that are built by your organization in different teams in Copilot Studio. So, you don’t have to remake and rebuild.” - Sarah Critchley

“We focus in on conversation management, user authorization management, state management. So, how do I build and expand my farm and, in turn, integrate with the rest of the Microsoft ecosystem.” - Matthew Barbour

Technical Architecture Deep Dive

Multi-Agent Communication Protocols

Agent Orchestration Layer:
??? Copilot Studio ? Copilot Studio (native protocol)
??? Copilot Studio ? Azure AI Foundry (connection string + API)
??? SDK ? Copilot Studio (client library integration)
??? Cross-platform ? A2A Protocol (future Google integration)
??? Event system ? Triggers and autonomous activation

Development Workflow Integration

Professional Development Stack:
??? Visual Studio Code Extension
?   ??? Clone/Push operations
?   ??? Local YAML editing
?   ??? IntelliSense and error detection
?   ??? Offline development capability
??? M365 Agents SDK
?   ??? C#, JavaScript, Python support
?   ??? Custom orchestration and models
?   ??? Multi-channel deployment
?   ??? Enterprise conversation management
??? Azure AI Foundry
    ??? Model hosting and management
    ??? Agent service integration
    ??? Cross-platform orchestration

Enterprise Integration Patterns

Organizational Development Models:
??? Bottom-up: Process owners using Copilot Studio
??? Top-down: Enterprise architects using SDK
??? Hybrid: Mixed development approaches
??? Integration: Cross-team agent collaboration

Implementation Guidelines

Choosing Development Approaches

**Copilot Studio (Low-Code):**

- Process owners and subject matter experts
- Rapid prototyping and visual development
- Single team focused on specific agents
- Agents that need direct end-user availability

**M365 Agents SDK (Pro-Code):**

- Professional developers and technical teams
- Custom control over orchestration and models
- Existing application integration
- Complex multi-agent architecture patterns
- Enterprise-scale deployment requirements

Multi-Agent Architecture Patterns

**Lightweight Specialized Agents:**

- Single maker or team development
- Logically grouped knowledge and tools
- Portable across environments
- Instruction-driven behavior definition

**Connected External Agents:**

- Multiple teams managing separate agents
- Direct end-user agent availability
- Cross-platform integration requirements
- Independent agent lifecycle management

Integration Best Practices

**Tool Sharing Strategy:**

- Configure tool availability at root agent level
- Use "Additional details" settings for usage control
- Separate shared tools from agent-specific tools
- Plan tool access patterns across agent hierarchy

**Conversation Context Management:**

- Choose full context vs. private mode based on security needs
- Plan conversation history flow across agent boundaries
- Implement appropriate authentication strategies
- Design for multi-turn conversation continuity

Advanced Applications and Use Cases

Enterprise Banking Transformation

Comprehensive Financial Services:

  • Account management - Balance inquiries, transaction history, account switching
  • Security operations - Lost/stolen card processing with conditional dispute handling
  • Loan services - Mortgage information, car loans, eligibility assessment
  • Cross-platform integration - Copilot Studio and Azure AI Foundry coordination

Global Enterprise Deployment

Multilingual Business Operations:

  • 29-language support - Global deployment with automatic localization
  • Cultural adaptation - Conversation patterns adapted to regional preferences
  • Centralized management - Single agent codebase serving multiple markets
  • Compliance integration - Regional regulatory requirements and data handling

Developer Productivity Enhancement

Professional Development Workflows:

  • Local development - Offline agent creation and testing
  • Version control - Git-like workflows for agent collaboration
  • IDE integration - Full IntelliSense and error detection for agent logic
  • Cross-platform deployment - Single codebase targeting multiple channels

Resources and Further Learning

Official Documentation and Tools

Architecture and Best Practices

Learning and Community

  • Enterprise Agent Challenge - May 28 - June 13 development competition
  • Breakout Session 163 - Advanced multi-agent SDK patterns and implementation
  • Build 2025 Labs and Demos - Hands-on learning opportunities across agent development

About the Speakers

Salem Bacha
Principal Architect
Microsoft
Architect focused on generative AI capabilities in Copilot Studio. Extensive experience in chatbot development and agent orchestration since joining Microsoft.

Gary Pretty
Principal Product Manager, Copilot Studio
Microsoft
Principal Product Manager driving generative orchestration capabilities and multi-agent features. Former Microsoft AI MVP with deep expertise in conversational AI platforms.

Sarah Critchley
Principal PM, Microsoft 365 Agents SDK
Microsoft
Principal Product Manager leading the M365 Agents SDK development and cross-platform integration capabilities.

Matthew Barbour
Principal Architect, Power Platform
Microsoft
Principal Architect and Development Manager for the Agents SDK, focused on developer ecosystem and enterprise integration patterns.

This comprehensive session demonstrates Microsoft’s vision for seamless multi-agent orchestration across development approaches, platforms, and organizational structures. The unified architecture enables organizations to build sophisticated agent ecosystems that transform business processes while providing choice in development methodology and deployment strategy.