ag2ai/ag2

What it solves

AG2 is a programming framework designed to simplify the creation of AI agents and the coordination of multiple agents to solve complex tasks. It provides a structured way to handle agent-to-agent communication, tool integration, and long-term memory, reducing the manual effort required to build agentic AI systems.

How it works

AG2 operates as an "AgentOS," utilizing a core Agent class that manages interactions with LLM providers and tool execution. It employs a Network system consisting of a Hub and typed channels (such as conversation, consulting, discussion, and workflow) to orchestrate how agents communicate. To enhance agent capabilities, it uses a harness that adds opt-in primitives like KnowledgeStore for persistent memory and compaction strategies (e.g., SummarizeCompact) to manage conversation history growth.

Who it’s for

It is intended for developers and researchers building agentic AI applications that require multi-agent cooperation, human-in-the-loop validation, and persistent knowledge across sessions.

Highlights

• Multi-Agent Orchestration: Uses a Hub-and-channel architecture to coordinate multiple agents via various communication patterns.
• Tool Integration: Allows plain Python functions to be converted into agent tools using a simple @tool decorator.
• Human-in-the-Loop: Built-in support for pausing agent runs to collect human confirmation or input via hitl_hook.
• Persistent Knowledge: Provides disk-backed knowledge stores and memory policies to ensure agents remember facts across different runs.
• Async-First Design: The framework is built to be asynchronous throughout for efficient agent interactions.