Core Architecture

Dapr Agents are intelligent building blocks that combine LLM reasoning with tool integration, memory, and collaboration features to create scalable agentic systems.

Key Features

LLM Integration & Outputs: Provides unified interfaces to connect with LLM APIs and leverages structured outputs following JSON Schema and OpenAPI standards for reliable, predictable results.

Dynamic Tool Selection: Agents automatically choose appropriate tools for tasks using LLM analysis and Function Calling capabilities, with built-in Model Context Protocol (MCP) support for discovering external tools at runtime.

Memory & Context: Agents maintain context across interactions through various memory options, from simple chat history to vector databases and Dapr state stores for persistent, scalable memory.

Service Architecture: Agents are deployed as independent FastAPI services with Dapr, enabling modular deployment and easy integration into multi-agent systems.

Agent Patterns

The document describes built-in patterns that define how agents operate:

• Tool Calling: Enables dynamic interaction with external tools through structured JSON outputs
• ReAct (Reason + Act): A cyclical pattern where agents think, act, and observe results to adapt and learn

Collaboration Framework

Agents collaborate through:

• Message-driven communication via Pub/Sub messaging for asynchronous, event-driven coordination
• Workflow orchestration supporting both deterministic and event-driven multi-agent workflows

Workflow Types

• Random Workflow: Randomly selects next agent for diversity in responses
• Round Robin: Sequential task assignment ensuring equal participation
• LLM-Based Workflow: Uses LLM reasoning to dynamically choose the most suitable agent based on context, history, and agent metadata

The framework emphasizes flexibility, modularity, and scalability for building sophisticated multi-agent AI systems.