#7070th of 142 in Agent frameworks and SDKs

ag2

Python framework for building single agents and multi-agent networks

Stars
5.1k
License
Apache-2.0
Last commit
Nov 2026
Last release
Nov 2026
Language
Python
Stars, last 29 weeks+682
22 Apr 2026: 4,378 stars29 Apr 2026: 4,411 stars6 May 2026: 4,467 stars13 May 2026: 4,485 stars20 May 2026: 4,536 stars27 May 2026: 4,559 stars3 Jun 2026: 4,587 stars10 Jun 2026: 4,617 stars17 Jun 2026: 4,641 stars24 Jun 2026: 4,669 stars1 Jul 2026: 4,688 stars8 Jul 2026: 4,710 stars15 Jul 2026: 4,741 stars22 Jul 2026: 4,752 stars29 Jul 2026: 4,779 stars5 Aug 2026: 4,801 stars12 Aug 2026: 4,822 stars19 Aug 2026: 4,847 stars26 Aug 2026: 4,861 stars2 Sept 2026: 4,873 stars9 Sept 2026: 4,889 stars16 Sept 2026: 4,911 stars23 Sept 2026: 4,932 stars30 Sept 2026: 4,952 stars7 Oct 2026: 4,969 stars14 Oct 2026: 4,986 stars21 Oct 2026: 5,002 stars28 Oct 2026: 5,025 stars4 Nov 2026: 5,042 stars11 Nov 2026: 5,060 stars

4,378 on 22 Apr 20265,060 on 11 Nov 2026

From GitHub's daily star counts.

Checked on GitHub.

Overview

AG2 v1.0 is an async Python library (import ag2) where an Agent calls a model provider, runs @tool functions and returns a reply. Multiple agents coordinate through a Network: a Hub with a write-ahead log and typed channels (conversation, consulting, discussion, workflow). Optional harness pieces add a knowledge store, memory injection and history compaction. The older autogen API now lives in ag2-classic.

Who it is for: Python developers building multi-agent LLM applications

Strengths

  • Typed Network channels with a hub write-ahead log and audit trail for multi-agent runs
  • Opt-in knowledge store and summarizing history compaction, with disk-backed persistence
  • Human-in-the-loop via context.input() inside tools, with a pluggable hitl_hook
  • Provider extras for OpenAI, Anthropic, Gemini and Ollama; minimal default dependencies

Weaknesses

  • v1.0 is not a drop-in upgrade from the autogen-based classic API
  • Requires Python 3.11 or newer
  • Async-only API; no synchronous entry point shown
  • Maintained by volunteers across organizations, per the README

What it needs

  • no GPU
  • Needs Python >= 3.11, LLM provider API key (OpenAI, Anthropic, Gemini) or Ollama
  • Models: OpenAI, Anthropic, Gemini, Ollama

What changed

  1. “A2UI v1.0 now speaks the stable upstream spec. 1.3.0 targeted a release candidate. v1.0 messages carry "version": "1.0" instead of "v1.0", the basic catalog requires TextField.value, and custom v1.0 …”

  2. “OpenAIConfig is deprecated and becomes the Responses API in 1.4.0. The current Chat Completions behavior is now OpenAICompletionsConfig; OpenAIConfig still works unchanged but emits a DeprecationWarn…”

From the release notes. No flag means the note does not say so, not that the upgrade is safe.

Also in Agent frameworks and SDKs

See all 142
Also in Agent frameworks and SDKs
RankProjectScore
1ppt-masterAgent skill that turns documents into natively editable PowerPoint decks Live demo ↗ (opens in a new tab)76.2k stars+15,031 (+25%) in 30 daysMIT96 out of 100
2openai-agents-pythonPython SDK for multi-agent workflows with handoffs, guardrails and tracing30.6k stars+692 (+2.3%) in 30 daysMIT92 out of 100
3mastraTypeScript framework for building AI agents, workflows and MCP servers29.5k stars+778 (+2.7%) in 30 dayscustom license92 out of 100
#4browser-usePython library that lets an LLM agent control a web browser Live demo ↗ (opens in a new tab)120.8k stars+3,334 (+2.8%) in 30 daysMIT90 out of 100
#5copilotkitReact and Angular SDK for adding agent chat and generative UI to apps38.7k stars+757 (+2.0%) in 30 daysMIT90 out of 100