Download Latest Version v0.13.0 source code.tar.gz (7.0 MB)
Email in envelope

Get an email when there's a new version of AG2

Home / v0.13.0
Name Modified Size InfoDownloads / Week
Parent folder
README.md 2026-05-13 8.9 kB
0
v0.13.0 source code.tar.gz 2026-05-13 7.0 MB
0
v0.13.0 source code.zip 2026-05-13 8.7 MB
0
Totals: 3 Items   15.7 MB 0

Highlights

🛣️ The Path to v1.0 — A Major Step

v0.13.0 is a significant release for our journey to V1.0. It introduces a brand-new Multi-Agent Network runtime and LiveAgent for full-duplex realtime voice. Plus, A2A v1.0 and MCP now in Beta to complete the connectivity story.

📍 Release Roadmap

🌐 Network — Multi-Agent Runtime (NEW)

Other frameworks treat an agent or a graph as the primitive. AG2 Beta makes the channel between agents the primitive — a durable, identity-scoped, protocol-typed container. Open it, work inside it, close it. The channel is the artifact.

  • 🔗 Channel as the unit – Append-only event log per channel. Adapter state is derived from the log, so a hub restart is a non-event.
  • 🎭 Choreography you can dial in – From open multi-party discussion to fully declared workflows, all on the same primitive. Four built-in patterns ship today:
    • consulting — 1↔1, one round (strict Q+R, auto-close)
    • conversation — 1↔1, multi-turn (bidirectional, long-running)
    • discussion — N-party turn-taking (round-robin, dynamic, or static ordering)
    • workflow — orchestrated by a declarative TransitionGraph, with conditions and dynamic handoffs (great for transitioning your classic AG2 group chats - Guidance)
  • 🛠️ Design your own orchestration – The ChannelAdapter protocol is small (~50 lines for a custom adapter). Define your own manifest, state, and turn-taking rules; reuse the durability and identity layers for free.
  • 🪪 Identity is three records – Passport (immutable, hub-stamped), Resume (mutable track record + observed stats), SKILL.md (LLM-facing doc). Discovery ranks by observed outcomes.
  • 🤝 Federation with Passport + Visa – A channel can span two hubs the same way a TCP connection spans two networks. Cross-org channels are the base case, not an extension.
  • 👤 HumanClient – Humans are peers on the network, not a special case.

🎮 Try it in the Playground – An 8-page hands-on walkthrough (~30 min) with full code examples. Build a hub, register agents, open channels, walk through all four turn-taking patterns, hub-side governance, and write your own adapter from scratch.

📔 Docs: Multi-Agent Network

🎙️ LiveAgent — Realtime Voice (NEW)

New autogen.beta.live module for full-duplex voice agents backed by provider realtime APIs. Unlike turn-by-turn STT/TTS pipelines, LiveAgent opens a single bidirectional session for the entire conversation — audio flows in and out continuously.

  • 🎤 Single bidirectional session – No turn-by-turn round trips.
  • 🛑 Built-in VAD and barge-in – Voice activity detection and interruption support out of the box.
  • 🛠️ Same tool surface as Agent – @agent.tool decorator works identically; tool calls route through AG2's normal executor.
  • 🌐 Provider-neutral – OpenAI Realtime (10 voices: alloy, ash, ballad, coral, echo, sage, shimmer, verse, marin, cedar) and Gemini Live (8 voices: Aoede, Charon, Fenrir, Kore, Leda, Orus, Puck, Zephyr).
  • 📝 Text-only mode – Keep the realtime session for low-latency turn detection but disable audio output entirely.
  • 🔊 Audio peers – SoundDevicePlayer and SoundDeviceRecorder share the conversation context with the agent.

📔 Docs: LiveAgent · STT & TTS alternative also available.

🔗 A2A v1.0 — Now in Beta

A2A is now available in the Beta framework via the new autogen.beta.a2a module.

  • 📡 Expose any AG2 Agent over A2A via A2AServer
  • 🔌 Connect to a remote A2A endpoint as if it were an LLM provider via A2AConfig
  • 🚚 Transport-agnostic: JSON-RPC, HTTP+JSON (REST), or gRPC
  • 📥 Tasks & push notifications for long-running work
  • Install with pip install "ag2[a2a]"

📔 Docs: A2A Overview

🔌 MCP — Now in Beta

Native MCP support in Beta with two connection patterns:

  • 🔄 Client-side (MCPServer) – AG2 connects to the MCP server, discovers tools, and executes them locally. Works with every provider. Supports remote HTTP and local stdio subprocesses.
  • 🚀 Provider-side (MCPServerTool) – Server URL and credentials forwarded to the LLM provider (e.g. Anthropic).

📔 Docs: MCP Servers

🔬 Other Beta Updates

  • 🧬 Task Lifecycle Primitive – New foundation for managing long-running agent tasks.
  • 🧠 Gemini/Vertex AI Thinking Level – Thinking Level support added to the Gemini/Vertex AI client with thinking token usage tracking.
  • ⚙️ Anthropic extra_body Config – New configuration support for passing extra_body to Anthropic requests.
  • 📚 Knowledge: Opt-Out Flags + Observable Aggregation Lifecycle – More control over harness behavior, plus observability into aggregation.
  • 🔧 Typed Gemini ToolCallEvent Subclass – Replaced ToolCallEvent.provider_data with a typed Gemini subclass.

🔐 Security

  • 🛡️ SSRF Protection in Document Agent – URL validation added to prevent Server-Side Request Forgery.

Bug Fixes

  • 🔧 Beta: Gemini Schema – Strip additionalProperties recursively to fix anyOf rejections.
  • 🔧 Beta: Policies – Pair tool calls/results correctly after history trim.
  • 🔧 LLMConfig in Notebooks – Fixed LLMConfig usage across 5 notebooks.
  • 🔧 uv sync Resolution – Aligned optional-extras pins so uv sync resolves.

Documentation & Maintenance

  • 📔 Switched interop-pydantic-ai extra to pydantic-ai-slim.
  • 🧪 Various test fixes and LLM provider/package alignments.

What's Changed

New Contributors

Full Changelog: https://github.com/ag2ai/ag2/compare/v0.12.3...v0.13.0

Source: README.md, updated 2026-05-13