Agent architecture components

Agent architecture components are the fundamental building blocks that comprise modern AI agent systems. These components work together to enable agents to receive user input, process requests, access external resources, and deliver meaningful responses through a coordinated system of interconnected services and interfaces.

Understanding these components is essential for architects and developers who design, implement, and maintain agent solutions within the Microsoft ecosystem. Each component serves a specific purpose in the agent workflow. The selection and configuration of these components directly affect the agent's capabilities, performance, and reliability.

The architectural decisions you make regarding these components determine whether an agent is suitable for simple conversational tasks or complex enterprise workflows that require integration with multiple systems and data sources.

Core architectural components

Most agents include a client interface, foundational infrastructure, an orchestrator, a language model, and tool calling capabilities. This standard architecture provides the framework for implementing agents across different Microsoft platforms.

• Client: The interface through which users interact with the agent, such as a chat window in Microsoft Teams or a custom application.
• Chat/app/storage infrastructure: Provides the foundational services for message handling, state management, and persistent storage of conversations or context.
• Orchestrator: Coordinates the flow of information between components. It manages tasks like intent recognition, workflow execution, and routing requests to the appropriate tools or services.
• Language model: The AI engine that interprets user input, generates responses, and powers natural language understanding and generation.
• Catalog: A registry or directory of available skills, plugins, or actions that the agent can invoke to extend its capabilities.
• Tool calling: The mechanism that enables the agent to invoke external functions, APIs, or services to fulfill user requests beyond simple conversation.
• OpenAPI based APIs: Standardized interfaces that allow the agent to connect with external systems and services using OpenAPI specifications.
• Semantic indexes: Structures that enable the agent to search, retrieve, and reason over enterprise data using semantic understanding rather than just keywords.
• MCP servers: Middleware that implements the Model Context Protocol, enabling agents to exchange context and capabilities with external tools and services securely.
• Proprietary protocols (such as Direct Line): Custom communication channels used to connect agents with specific platforms or services, such as Direct Line for Bot Framework.
• Parsing and chunking: Processes that break down user input or documents into manageable pieces for more effective analysis and response generation.
• RAI (Responsible AI): Ensures ethical and compliant behavior across the agent lifecycle, including fairness, transparency, and privacy safeguards.

Orchestrators and language models

Orchestrators are components that organize and route prompts to the correct subcomponent. They can include AI or language model-enhanced flows (for example, generative orchestration in Copilot Studio), or deterministic code flows (like topics in Copilot Studio, commonly used for task management such as audit logging or catalog retrieval). SDKs like Semantic Kernel or LangChain allow for both workflow-based and language model-based orchestration at the cost of increased complexity of the overall solution.

When gathering requirements for an agentic solution, gather sufficient information about the scenario to determine if dynamic intent assessment and task execution adds value (language model orchestrator), or if a deterministic, highly accurate and precise workflow with no tolerance for variation is needed (code-based orchestrator).

Catalogs

The Unified app catalog for Microsoft 365 provides a managed, governed agent catalog that allows users to discover, acquire, and manage custom engine agents and declarative agents for Microsoft 365 Copilot surfaces.

External catalogs exist for scenarios that involve cross agent platform acquisition, like Model Context Protocol. These catalogs can complement the Microsoft 365 catalog when published to the Unified App catalog within a Microsoft tenant.