microsoft/agent-framework

Features

• Autonomous AI Agent Frameworks - Provides a comprehensive framework for building and deploying autonomous AI agents that use language models.
• Multi-Agent Orchestration - Coordinates interactions between multiple AI entities using sequential or concurrent patterns to achieve complex goals.
• Multi-Agent Coordination Frameworks - Coordinates the execution of multiple agents using directed graphs to manage data flow and task orchestration.
• Long-term Memory Stores - Implements persistent storage mechanisms to retain user context and memories across multiple agent sessions.
• Agent Communication Protocols - Implements common protocols to enable discovery and message exchange between agents across different frameworks.
• Agent Provider Integrations - Integrates agents with various AI model providers and specialized processing services.
• Agent Session Management - Saves and reloads full session state objects to resume conversations and maintain continuity across restarts.
• Agent Workflow Orchestrations - Sequences specialized AI agents and functions via graph-based paths with type-safe routing.
• Autonomous Agents - Builds autonomous entities that integrate LLMs with tool usage and planning capabilities to perform tasks.
• Agent Orchestration Frameworks - Acts as an orchestration framework for building autonomous agents that coordinate multi-agent workflows.
• Agent Instantiation - Enables building autonomous entities that process requests using a small set of configuration.
• Agent Construction Frameworks - Provides a framework for defining how custom data, memory, and tools are integrated into autonomous agents.
• Stateful Workflow Persistence - Saves progress of long-running processes via state checkpoints to enable recovery from the last known state.
• Durable Agent Runtimes - Ensures workload durability by persisting agent and workflow state to recover long-running tasks after failures.
• Multi-Agent Orchestration Patterns - Uses deterministic code-based workflows to manage complex interactions between agents in sequential or parallel patterns.
• Agent Response Streamers - Ships a mechanism for streaming incremental agent outputs and tool interactions to clients in real time.
• Agent Tool Integrations - Connects agents to external APIs, database queries, and functions to execute specific tasks automatically.
• Agentic Workflow Graphs - Models task sequences as directed graphs to support conditional routing and parallel execution of autonomous agents.
• AI Agent Orchestration Frameworks - Implements runtimes that manage agent loops and interact with underlying inference services via direct configuration.
• LLM Tooling Integrations - Provides a dedicated integration layer for binding external functions and APIs as executable tools for LLMs.
• Conversation History Management - Stores and manages logs of agent-user interactions to maintain multi-turn context.
• Conversation State Management - Implements conversational state management to persist chat history and session memory across interactions.
• Conversational Session Management - Maintains conversation history across multi-turn dialogues by mapping unique identifiers to session stores.
• Conversational State Managers - Provides a state persistence layer that tracks session history and conversation threads across various stores.
• Human Approval Gates - Provides human-in-the-loop approval gates to authorize sensitive tool invocations before the model processes the results.
• Persistent Chat Histories - Persists conversation records using NoSQL or key-value stores to maintain long-term history across sessions.
• Conversation Threads - Provides persistent containers for organizing user-agent message history as conversation threads.
• LLM Provider Integrations - Provides connectivity and configuration for various inference services including cloud APIs and local open-source models.
• Retrieval-Augmented Generation Frameworks - Includes a framework for executing retrieval-augmented generation pipelines to ground model responses in external data.
• Stateful Run Executions - Implements the runtime execution of agents while maintaining conversation state across sequential steps.
• Human-in-the-Loop Approvals - Implements mechanisms to pause autonomous execution and require human approval before tool invocation.
• Sequential Step Orchestrators - Links multiple processing steps in a sequential chain where data flows from one executor to the next.
• Vector Database Abstractions - Provides a unified abstraction layer to connect agents to multiple vector database backends for efficient data retrieval.
• Message Routing - Directs workflow message flow using conditional branching and parallel fan-out patterns.
• State Checkpointing - Records the execution state of graphs to avoid repeating completed steps after a system failure.
• In-Process Message Routing - Implements internal message routing between processing units to coordinate data flow within the agent framework.
• Control Flow Logic Models - Defines execution paths via directed graphs or functional logic to implement conditional routing.
• Graph-Based Workflow Models - Uses graph architectures to model complex task sequences supporting parallel processing and conditional routing.
• Workflow Logic Engines - Implements the underlying execution logic and graph edges required to coordinate complex agent task sequences.
• Typed Message Processing - Executes custom logic by receiving and processing typed messages within dedicated workflow processing units.
• Agent Connectivity Interfaces - Implements protocols for resolving agent capabilities via metadata and establishing communication across remote endpoints.
• Agent Definitions - Allows defining specialized agent behaviors and roles by extending core base abstractions.
• Agent Function Libraries - Provides agents with libraries of specialized external functions to perform tasks and access real-time data.
• Session Personalization - Injects personalized user information into agent runs via context providers to tailor behavior.
• Agent Skill Frameworks - Provides a flexible system for loading agent capabilities from files, inline code, or classes.
• Agent State Persistence - Saves and restores the execution state of autonomous agents to ensure continuity after failures or restarts.
• Agent Capability Extensions - Provides reusable modules that package domain expertise to expand the functional capabilities of agents.
• Agent System Prompts - Allows customization of system prompts to control how agent skills and tools are presented to the model.
• Agent Configurations - Uses typed dictionaries to define and manage inference parameters and provider settings for agents.
• Conversational Agent Construction - Supports the construction of agents for multi-turn conversations with managed history and structured outputs.
• Agent Tooling Definitions - Provides mechanisms to define tool signatures that allow agents to request execution on remote systems.
• Domain Expertise Packages - Provides modular domain expertise packages that agents load to perform specialized professional tasks.
• OpenAI-Compatible APIs - Exposes a backend API that follows the OpenAI specification for compatibility with standard AI SDKs.
• Agent Message Proxies - Enables connection and invocation of agents hosted on remote services through standardized message proxying.
• Workflow Validation - Verifies graph connectivity and type compatibility to ensure the execution graph is logically sound.
• Runtime Context Injections - Passes session-specific data to tool functions without exposing internal parameters to the model's schema.
• Context Window Optimizations - Optimizes token usage by implementing high-level summaries and on-demand loading of detailed instructions.
• Session History Retrieval - Reconstructs conversation state and retrieves historical data using service identifiers for context restoration.
• Custom Provider Implementations - Provides a base class to implement custom agent types for proprietary or unsupported inference services.
• External Service Integrations - Connects workflows to external APIs and integrates human-in-the-loop interaction patterns.
• External System Integrations - Integrates agent logic with durable extensions, communication protocols, and developer interfaces.
• External Tool Execution - Executes function tools, code interpreters, and file search capabilities provided by the inference service.
• Function-to-Tool Converters - Converts standard code methods into executable agent tools with automatic schema generation.
• Hierarchical Agent Orchestration - Supports hierarchical agent structures where one agent can be wrapped as a tool for another.
• Human-in-the-Loop Workflows - Incorporates human-in-the-loop workflows with approval gates to authorize tool use or review agent actions.
• Inference Configuration Parameters - Sets provider-specific inference parameters, such as output tokens, using typed configuration dictionaries.
• Model Provider Abstractions - Decouples agent logic from specific LLM providers using a base class abstraction layer.
• Schema Enforcement Layers - Constrains model outputs to adhere to specific JSON schemas during the generation process.
• Multi-Agent Orchestrators - Coordinates teams of specialized agents to automate complex business processes via sequential or concurrent patterns.
• Skill Availability Controls - Allows adding or restricting available tools during an active session based on the current agent state.
• Structured Output Converters - Uses a transformation layer to convert plain text model responses into validated JSON schemas.
• Tool Calling - Enables agents to iteratively execute external functions until a task is complete, including loop safety mechanisms.
• Tool Schema Definitions - Implements standardized formats for function names and parameter types to guide model tool selection.
• Streaming Workflow Execution - Streams final answers and intermediate progress updates back to the caller during workflow execution.
• Session State Management - Persists conversation history and session state using external stores such as Redis.
• Text-to-JSON Converters - Transforms plain text responses into structured JSON using a decorator pattern for models without native support.
• Human-in-the-loop Interfaces - Suspends execution to wait for external events or human approval without consuming active compute resources.
• Tool Function Registrations - Binds functions directly to an agent during creation or execution to extend capabilities without specialized attributes.
• Session State Serializers - Serializes conversation data and agent states to allow restoration from stored snapshots.
• Middleware-Based Request Pipelines - Provides a modular chain of handlers to intercept agent operations for logging, safety filtering, and telemetry.
• Agent Runtime Exposure - Wraps internal agents in a standardized server-side protocol for discovery and invocation by external clients.
• Agent Telemetry Streams - Implements real-time telemetry streams to track agent logs, request metadata, and token consumption.
• Agent Endpoint Access Control - Intercepts requests to agent endpoints to verify identities and enforce access control policies.
• Session Identifiers - Uses unique session identifiers to reconnect agents to existing conversations and reload state.
• Agentic Session Persistence - Tracks task progress and session state in external stores to allow agents to resume work across sessions.
• Action Dispatch Middleware - Applies middleware to agent operations to implement telemetry, safety mitigations, and business logic.
• Background Task Management - Executes processes in the background and provides continuation tokens to poll for results or subscribe to updates.
• Dependency Injection Containers - Ships a central service registry to manage shared resources, tools, and agent configurations via dependency injection.
• Dependency Injection - Injects service providers into skill resources to allow agents access to external data and business logic.
• Operation Interceptors - Intercepts internal agent operations to implement input validation, content filtering, logging, and caching.
• Request Middleware - Implements a request middleware pipeline to handle cross-cutting concerns like security filtering and rate limiting.
• Agent Execution Tracing - Ships a dashboard to visualize agent reasoning trajectories and tool usage for performance analysis.
• Agent Observability - Includes telemetry and tracing capabilities to monitor agent performance and debug complex multi-step workflows.
• Agent Performance Monitoring - Tracks operational metrics and distributed tracing data to observe and optimize agent behavior.
• Event Monitoring Systems - Tracks workflow progress through a system of lifecycle events and real-time state broadcasting.
• AI and Agent Observability - Ships an observability suite using OpenTelemetry to trace execution flows and monitor tool calls.
• GenAI Execution Monitoring - Provides specialized observability for tracking the internal execution flow and runtime behavior of generative AI components.
• Observability Instrumentation - Instruments AI workflows via environment variables and configuration to enable distributed tracing and telemetry.
• Agent State Tracking - Tracks session-based memory and internal state transitions for long-running tasks and human interactions.
• Agent Frameworks - Unified event-driven actor model for multi-agent orchestration.
• AI Agent Frameworks - Framework for orchestrating multi-agent workflows.

Star history