Architecting Multi-Agent Solutions with Copilot Studio and M365 Agents SDK
Session Date: May 20, 2025
Duration: 1 hour
Venue: Microsoft Build 2025 Conference - BRK176
Speakers: Vid Chari (Copilot Studio Marketing), Gary Pretty (Principal Product Manager, Copilot Studio), Salem Bacha (Principal Architect, Copilot Studio), Mike Stall (Architect, Copilot Studio), Sarah Critchley (Principal PM, Microsoft 365 Agents SDK), Matthew Barbour (Principal Architect, Power Platform & Development Manager, Agents SDK)
Link: Microsoft Build 2025 Session BRK176
Table of Contents
- The Multi-Agent Vision
- Microsoft’s Unified Platform
- Copilot Studio Multi-Agent Architecture
- Live Demonstrations
- Multilingual Capabilities
- VS Code Extension
- M365 Agents SDK
- Advanced Features
- Integration Patterns
- Future Roadmap
- References
1. The Multi-Agent Vision
Timeframe: 00:00:00 - 00:08:30 (8m 30s)
Speaker: Vid Chari
Vid Chari established Microsoft’s fundamental distinction between Copilots and Agents, addressing common market confusion.
Core Framework
Copilots:
- Personal AI assistants aligned to individual employees
- Grounded in personal data (emails, chats, meetings, documents)
- Universal “UI to AI” for Microsoft experiences
- One-to-one relationship with users
Agents:
- Programs using AI to automate business processes
- Aligned to organizational workflows and processes
- Range from simple retrieval to complex autonomous systems
- Work alongside individuals, teams, or organizations
Market Scale Prediction
IDC predicts 1.3 billion agents by 2028, driven by diverse organizational stakeholders:
- Bottom-up development: Process owners transforming specific business areas
- Top-down strategy: Enterprise initiatives enabling new business models
- Grassroots innovation: Individual contributors solving workflow inefficiencies
Microsoft Platform Advantages
- Direct UI Integration: All agents integrate with M365 Copilot interface
- Data Ecosystem: Most enterprise data already in Microsoft platforms
- Enterprise Security: Built-in governance, compliance, and security
- AI Access: Direct integration with latest models and orchestration
2. Microsoft’s Unified Platform
Timeframe: 00:08:30 - 00:12:15 (3m 45s)
Speaker: Vid Chari
Microsoft provides choice across the development spectrum with unified underpinnings.
Development Approaches
Low-Code: Copilot Studio
- Audience: Process owners, subject matter experts, citizen developers
- Features: Visual canvas, rapid authoring, pre-built defaults
- Use Cases: Process-specific transformations, rapid prototyping
Pro-Code: Visual Studio + GitHub
- Audience: Professional developers, technical teams
- Features: Custom control, enterprise-grade capabilities
- Use Cases: Organization-wide initiatives, complex integrations
Common Foundation
Both approaches share:
- Knowledge systems: Unified data grounding and retrieval
- Tools and connectors: Shared integration library
- AI models: Common Azure AI Foundry access
- Orchestration: Unified multi-agent coordination
3. Copilot Studio Multi-Agent Architecture
Timeframe: 00:12:15 - 00:18:30 (6m 15s)
Speakers: Gary Pretty, Salem Bacha
Traditional Architecture
Current agent components:
- Generative Orchestrator: Recently GA worldwide
- Knowledge Sources: Document grounding and retrieval
- Tools and Actions: External system integration
- Topics: High-control workflow definitions
3.1 Lightweight Specialized Agents
Design Philosophy:
- Embedded within parent agents
- Instruction-driven behavior
- Focused task specialization
- Environment portability
Key Capabilities:
- Natural Language Instructions: Conversational behavior definition
- Tool Inheritance: Granular access to parent tools
- Built-in Interaction Tools: Pre-built question/message capabilities
- Environment Portability: Move with parent agent
3.2 Connected External Agents
Architecture Benefits:
- Independent lifecycle management
- Cross-team development and ownership
- Direct end-user availability
- Platform-agnostic integration
Integration Features:
- Conversation History Control: Full context vs. private mode
- Authentication Options: User vs. service account credentials
- Multi-turn Support: Seamless conversation continuity
- Cross-platform Protocol: Unified communication
4. Live Demonstrations
Timeframe: 00:18:30 - 00:35:45 (17m 15s)
Speakers: Gary Pretty, Salem Bacha
4.1 Balance Information Agent
Timeframe: 00:18:30 - 00:23:00 (4m 30s)
Speaker: Gary Pretty
Configuration:
- Name: Balance Information Agent
- Description: Provides balance information for user accounts
- Instructions: Three-step workflow for account balance retrieval
Demo Results:
- Dynamic questioning when account not specified
- Seamless tool integration with shared resources
- Natural follow-up question handling
- Complete conversation generation without predefined inputs
4.2 Lost/Stolen Card Workflow
Timeframe: 00:23:00 - 00:27:30 (4m 30s)
Speaker: Gary Pretty
Complex Instructions:
- Get user accounts list and request selection
- Format account list with presentation rules
- Freeze account and retrieve transactions
- Show five recent transactions, check for suspicious activity
- If suspicious found, call dispute endpoint
Advanced Features:
- Multi-tool orchestration between user interactions
- LLM-driven conditional processing
- Natural language input handling
- Complete execution transparency
4.3 Cross-Platform Integration
Timeframe: 00:27:30 - 00:35:45 (8m 15s)
Speaker: Salem Bacha
Copilot Studio to Copilot Studio
Mortgage Agent Connection:
- Simple toggle-based availability
- Publishing requirement for connectivity
- Customizable interaction descriptions
- Full context vs. private mode options
Azure AI Foundry Integration
Car Loan Agent Setup:
- GPT-4.1 model integration
- Connection via URL, ID, and connection string
- Authentication options (user vs. author)
- Multi-turn conversation support
Results: Seamless cross-platform agent orchestration with conversation threads visible in both platforms.
5. Multilingual Capabilities
Timeframe: 00:35:45 - 00:42:00 (6m 15s)
Speaker: Gary Pretty
Revolutionary Announcement
Public Preview: Additional language support for generative orchestrator across all 29 Copilot Studio languages.
Key Features:
- Single Development Language: Build agents entirely in English
- Automatic Translation: Real-time conversation translation
- Zero Additional Work: Enable by checking language boxes
- Weeks to GA: General availability within weeks
Live Spanish Demonstration
Technical Achievement:
- Agent built in English operated flawlessly in Spanish
- Complex lost card workflow conducted entirely in Spanish
- Natural language responses throughout conditional logic
- Single codebase serving multiple global markets
Impact
- Historical Context: Previous implementations required extensive manual translation
- Business Impact: Rapid international expansion capability
- Development Efficiency: 29-language deployment with single-language effort
6. VS Code Extension
Timeframe: 00:42:00 - 00:50:30 (8m 30s)
Speaker: Mike Stall
Capabilities
Now Available: Public preview in VS Code marketplace with daily updates.
Core Features:
- Clone/Push Workflow: Git-like cloud synchronization
- Offline Development: Complete local development capability
- Rich IDE Support: IntelliSense, error detection, go-to-definition
- File-based Editing: YAML with semantic understanding
- Source Control Integration: Change tracking and collaboration
Development Workflow
Process:
- Clone agent from Copilot Studio URL
- Local YAML file development with full IDE support
- Offline capability for uninterrupted development
- Push changes back to cloud with live updates
Advanced Features:
- Semantic understanding of agent structure
- Color-coded change indicators
- Meaningful error messages with corrections
- Complete agent component navigation
Multilingual Solution
System Topic Challenge: Solved hardcoded welcome message limitation by replacing system topics with instruction-driven agents that inherit multilingual capabilities.
7. M365 Agents SDK
Timeframe: 00:50:30 - 01:00:00 (9m 30s)
Speakers: Sarah Critchley, Matthew Barbour
7.1 SDK Architecture
Speaker: Sarah Critchley
Definition: Open source SDK in C#, JavaScript, and Python providing complete development control.
Developer Control Areas:
- AI Models: Any service or provider choice
- Orchestrator: Custom logic and decision-making
- Knowledge Sources: Custom grounding and data
- Conversation Management: Built-in state, storage, auth
Multi-Channel Deployment:
- Microsoft Teams, Slack, M365 Copilot
- 15+ channels out-of-the-box
- Custom channel development capability
7.2 Technical Implementation
Speaker: Matthew Barbour
SDK Focus: Integrate existing custom agents with Microsoft ecosystem.
Core Capabilities:
- Channel Management: Abstract communication complexity
- Conversation Management: Multi-turn conversations and state
- User Authorization: Built-in authentication management
- State Management: Persistent session handling
Copilot Studio Client
Implementation:
var client = new CopilotStudioClient();
await client.StartAsync();
await client.StartConversationAsync();
await client.AskAsync("Hi, who are you?");Configuration Options:
- Direct connection URL for simple integration
- Environment ID for multi-agent switching
- Schema name for dynamic discovery
- Authentication flexibility
Advanced Integration:
- Semantic Kernel embedding (Python complete, .NET in progress)
- Full event infrastructure access
- Adaptive card and complex event support
- Platform-agnostic API consistency
8. Advanced Features
Trigger System
Event-Driven Activation:
- Conversation Start: User session initiation
- Orchestrator Decision: AI-driven agent selection
- External Events: System-triggered operations
- Custom Triggers: Business-specific conditions
Autonomous Operations
Both conversational and autonomous scenarios supported with identical multi-agent capabilities.
Activity Monitoring
Enhanced visibility includes:
- Agent invocation tracking
- Internal tool usage monitoring
- Cross-agent communication summaries
- Complete execution transparency
9. Integration Patterns
Development Approach Selection
Copilot Studio Scenarios:
- Process owner-led development
- Single team focused projects
- Rapid prototyping needs
- Direct end-user availability requirements
SDK Scenarios:
- Professional developer requirements
- Enterprise-scale implementations
- Existing application integration
- Complex orchestration patterns
Architecture Patterns
Lightweight Agents:
- Single development team scenarios
- Logically grouped knowledge and tools
- Portable with parent agent
- Instruction-driven behavior
Connected Agents:
- Multi-team management
- Independent end-user availability
- Cross-platform integration
- Independent lifecycle management
10. Future Roadmap
Immediate Availability (Build 2025)
- VS Code extension public preview
- Multilingual support (29 languages)
- Daily extension updates
June 2025 Public Preview
- Multi-agent orchestration
- Cross-platform coordination
- Complete business process transformation
Planned Expansions
- Google A2A protocol integration
- Extended cross-platform communication
- Enhanced SDK capabilities
- Expanded channel support
References
Official Documentation
Copilot Studio Documentation
Comprehensive guide to Microsoft’s low-code agent platform. Essential for understanding visual development approaches demonstrated in the session.Microsoft 365 Agents SDK
Open source SDK repository with complete source code, samples, and documentation for professional agent development.Azure AI Foundry Documentation
Platform documentation for AI infrastructure powering both Copilot Studio and SDK agents. Critical for understanding multi-agent orchestration foundation.
Development Tools
VS Code Copilot Studio Extension
Newly announced extension for professional developers. Enables familiar IDE tools for agent development with low-code platform integration.Semantic Kernel Documentation
Microsoft’s orchestration framework integrating with M365 Agents SDK. Relevant for complex multi-agent scenarios and custom orchestration.
Architecture Resources
Power Platform Guidance
Official architecture patterns and best practices for Power Platform. Valuable for enterprise multi-agent solution deployment strategies.Microsoft Power Platform Architecture
Enterprise architecture guidance for Power Platform implementations. Critical for large-scale agent deployments across multiple teams.Copilot Studio Starter Kit
Templates, best practices, and accelerators for Copilot Studio development. Useful for implementing demonstrated multi-agent patterns.
Industry Context
IDC AI Agents Market Forecast
Industry analysis supporting the 1.3 billion agents by 2028 prediction. Important for understanding business context driving Microsoft’s strategy.Microsoft Build 2025 Keynotes
Satya Nadella and Charles Lamanna keynotes referenced throughout the session. Provides broader context for Microsoft’s AI and agent announcements.
Additional Sessions
BRK163: Advanced Multi-Agent SDK Patterns
Follow-up session for advanced SDK implementation patterns and multi-agent switching scenarios.Enterprise Agent Challenge
May 28 - June 13 development competition for hands-on experience with demonstrated multi-agent capabilities.
Technical Samples
M365 Agents SDK Samples
Code samples including the Copilot Studio client console application demonstrated by Matthew Barbour.Azure AI Foundry Agent Templates
Templates for building Foundry agents that integrate with Copilot Studio agents in cross-platform scenarios.
This comprehensive analysis captures Microsoft’s revolutionary multi-agent capabilities announced at Build 2025, demonstrating unified orchestration across development approaches, platforms, and organizational boundaries while maintaining consistent user experience and enterprise-grade security.