AG2

The Open-Source AgentOS

[AG2 @ag2oss]

Most enterprises don’t have an AI strategy problem, they have an AI execution problem.

Even with proven ROI from successful pilots, agents often remain fragmented and unable to integrate with core business processes.

That changes today with https://t.co/j8VeJuRJBB. Powered by AgentOS, we transform fragmented experiments into a coordinated, production-ready AI workforce.

The winners of 2026 won't be the companies with the most AI pilots, they'll be the ones with the most scalable orchestration.

Ready to build your AI workforce? 👉 https://t.co/AUMPEhpCfY

#EnterpriseAI #AgenticAI #OpenSource

[AG2 @ag2oss]

AG2 v0.13.2 is here! 🛣️

🔬 Beta framework
🧪 NEW: Agent Evaluations + threshold() scorer for pass/fail gating — try it live at https://t.co/ltNpL7W8u0
📁 xAI File API support

⚙️ Classic framework
🤖 NEW: Anthropic V2 + Bedrock V2 clients (UnifiedResponse architecture)
🛡️ Security: CVE fix in ContextExpression (GHSA-9fvw-gr53-m7fw)
🚫 Shell operators blocked in readonly mode
🔧 Tool registration, DeepResearchAgent, LlamaIndex 0.13, MCP StdioConfig fixes

🐟 Framework-wide: TinyFish Search & Fetch tools (Beta extension + Classic experimental)

🙌 Thanks to our contributors, including 4 first-timers!

Release: https://t.co/pC0mgS3VZp

#AG2 #AI #MultiAgentSystems #AIAgents #OpenSource

[AG2 @ag2oss]

AG2 v0.13.1 is here! 🛣️

🔬 Beta framework (autogen.beta)
⚡ New: xAI (Grok) provider
🙌 Dynamic Agents (agent creating agents on-the-fly)
🌐 Network: named views, workflow finish tool, consistent envelope rendering
🖥️ AG-UI: reasoning messages both ways + multimodal user content
🚦 ConditionalMiddleware + dynamic_agent tool factory

⚙️ Classic framework
🔧 Fixes: RemoteAgent (A2A), fast_depends + **kwargs tools, Responses image costs, async LangChain tools

🔁 Framework-wide: OpenAI SDK 2.x migration

🙌 Thanks to our contributors!

Release: https://t.co/uM8AbMalCA

#AG2 #AI #MultiAgentSystems #AIAgents #OpenSource

[AG2 @ag2oss]

The AG2 Network is a multi-agent runtime where the channel between agents holds the conversation state. So what happens when the hub process restarts?

The conversation survives exactly - envelope for envelope, same order, same state.

🔄 This post covers:
✅ The write-ahead log: append to disk first, fan out second
✅ Hub restart: agents reconnect to the same channel and continue
✅ Three identity records: Passport, Resume (ranked by real track record), SKILL.md
✅ The audit log: append-only record of everything hub-wide

Same durability contract as Postgres, Kafka, etcd.

See what survives 👇
https://t.co/OspVasXz9C

#AG2 #AI #MultiAgentSystems #AIAgents

[AG2 @ag2oss]

🎭 Part 2 of our Network series walks through three dials is now available , each with runnable code:
✅ Time: built-in expectations (acks_within, reply_within, max_silence)
✅ Visibility: audience addressing for private side-channels
✅ Routing: ContextEquals, ToolCalled, dynamic Handoff returns

Find the right orchestration and control for your agentic workflows.

Get stuck in 👇
https://t.co/qEzdw7NX3M

#AG2 #AI #MultiAgentSystems #AIAgents

[AG2 @ag2oss]

The new AG2 Network released beta is built around one design choice: the channel between agents is the primitive. Not the agent, not the graph.

🌐 The channel holds the state. The hub is a thin stateless router; restart it, every channel comes back from a persisted state.

🎭 Four conversation shapes ship today, each with runnable code:
✅ consulting: 1↔1, one round
✅ conversation: 1↔1, free-form
✅ discussion: N-party round-robin
✅ workflow: declarative graph with handoffs

Get stuck in 👇
https://t.co/NonYRfq5DU

#AG2 #AI #MultiAgentSystems #AIAgents

[AG2 @ag2oss]

With OpenAI's GPT Realtime 2 and Thinking Machines' interaction models research preview, natural human-to-agent voice is being realised.

On our path to v1.0, we've introduced realtime voice into AG2 Beta — ready for human-to-multi-agent voice interactions.

3 lines of code → a full-duplex voice agent with GPT Realtime 2 🎙️

AG2 Beta's LiveAgent:
✅ One bidirectional realtime session
✅ Built-in VAD + barge-in
✅ Same @agent.tool as a text Agent
✅ OpenAI + Gemini providers

Blog with code samples 👇
https://t.co/wc38Qq1LPC

#AG2 #AI #VoiceAI #AIAgents #Realtime

[AG2 @ag2oss]

AG2 v0.13.0 — a mega release! 🚀

🌐 NEW: Multi-Agent Network — channels (not agents) are the unit. Four built-in choreographies (consulting, conversation, discussion, workflow) + bring your own orchestration. Federation with Passport+Visa.
🎙️ NEW: LiveAgent — full-duplex realtime voice with VAD + barge-in (OpenAI + Gemini)

Plus, A2A v1.0 and MCP arrive in Beta.

🎮 Journey through Network now in our Playground: https://t.co/LDx8nIahjz

Release: https://t.co/4WGSVYAAoc

🙌 Thanks to our contributors!

#AG2 #AI #MultiAgentSystems #AIAgents #OpenSource

[AG2 @ag2oss]

AG2 v0.12.3 is here! 🛣️

🤝 A2A v1.0 spec compatibility — Remote agents with Agent2Agent
🏖️ Beta: SandboxCodeTool — safe code execution in Daytona or Docker
🔍 Beta: Search tools — PerplexitySearchToolkit, Exa
🔧 OpenAI image handling + Gemini fixes
📔 V1.0 Contribution Policy

Release: https://t.co/K1g3GwsOkT
Roadmap: https://t.co/JirvWGFf4b

#AG2 #AI #MultiAgentSystems #AIAgents #OpenSource

[AG2 @ag2oss]

AG2 v0.12.2 is here! 🏗️

The Agent Harness lands in Beta — turning stateless LLM loops into stateful agents that:
🧠 Remember across sessions
📁 Organize their own knowledge (virtual filesystem)
✂️ Manage their own context window
🤝 Share findings with peers

Plus:
🔍 PerplexitySearchTool in Beta
🔑 Workload Identity for OpenAI
🔧 Gemini, Redis & AG-UI fixes

Release: https://t.co/jwoyv2PAxz
New Beta Playground examples: https://t.co/ltNpL7W8u0
Roadmap: https://t.co/JirvWGFf4b

#AG2 #AI #MultiAgentSystems #AIAgents #OpenSource

[AG2 @ag2oss]

Excited to have @daytonaio join in supporting the AG2 Hackathon 🚀

They’ve added $3,000 in credits to the prize pool for the best project.

Build and ship real multi-agent systems in a day.

May 3 - NYC

Free + food 🍕

Spaces are limited: https://t.co/ZTshT1vixv

[AG2 @ag2oss]

AG2 v0.12.1 is here! 🛣️

👥 Community feedback win: PerplexitySearchTool stays (no longer deprecated)!
☁️ Beta: Google Vertex AI support
📁 Beta: Files API client, multi-part ToolResult, Toolkit merging
🔍 New search tools: Tavily, Exa, DDGS
🎯 Advanced orchestration with step events
📦 Docker moved to extra

🙌 Thanks to our contributors!

Release: https://t.co/LkI8XeQ6wM
Roadmap: https://t.co/JirvWGFf4b

#AG2 #AI #MultiAgentSystems #AIAgents #OpenSource

[AG2 @ag2oss]

Join us for our AG2 Hackathon in NYC on May 3!

Build and present a multi-agent AI project in one day using AG2 Beta - our brand new next-gen multi-agent framework.

Hear from IBM Fellow Dinesh Verma, join a hands-on AG2 workshop, then build alongside top engineers and founders.

🏆Prizes for best builds + AG2 credits, swag, and your work featured across the AI community.

Free to attend. Food included 🍕.

Limited spots - register now: https://t.co/9C44R9sniD

Can't wait to see the next gen AI agent builders in action 🚀

[AG2 @ag2oss]

AG2 v0.12.0 – The Path to v1.0 begins! 🛣️

🚀 Beta → v1.0 journey kicks off
⚠️ Deprecation notices (removed in v0.14) — community feedback window open!
🔬 Beta: Toolkits, Tasks, Observer API, AG-UI, Multimodal input, AgentSpec
🧰 New tools: FilesystemToolset, LocalShellTool, Skills
🎉 DaytonaCodeExecutor + NLIP
🛡️ CVE-2025-69872 fixed

Roadmap: https://t.co/JirvWGFf4b
Release: https://t.co/znffr33qVQ

🙌 Share your feedback!

#AG2 #AI #MultiAgentSystems #AIAgents #OpenSource

[AG2 @ag2oss]

93% of API teams hit integration walls (Postman 2025).

That's why we partnered with https://t.co/SgY4PTNZNQ: AG2 agents now connect to 8,000+ tools with zero credential management. One function call. OAuth handled.

Gmail example live: https://t.co/sGNBG8hUAU

#AG2 #AI #Arcade #OAuth #MCP #Gmail

[AG2 @ag2oss]

Your agents can reason and act. But can they express?

Most agents are limited to conversational responses. Users need cards, forms, dashboards, and interactivity.

Google's A2UI gives agents a visual vocabulary. AG2's new A2UIAgent generates rich, interactive UI — rendered natively on any platform.

One agent. Every surface.

https://t.co/eVHTaX6iKV

[AG2 @ag2oss]

AG2 v0.11.5 is here! 🚀

🖥️ AG2 CLI – build, run, and test multi-agent applications agents directly from the terminal. Empowers your coding assistant!

🖥️ A2UIAgent – A2UI for agentically built and dynamic frontends

🔬 Beta advances: Telemetry, ResponseSchema, RedisStream, WebFetchTool, tool-scoped middleware, MemoryTool, Image Generation, Shell Tool

📔 https://t.co/SgY4PTNZNQ integration blog + release roadmap

🙌 Thanks to our contributors!

Release: https://t.co/YHKnBzv5XF

#AG2 #AI #MultiAgentSystems #AIAgents #OpenSource #A2UI #CLI

[AG2 @ag2oss]

AG2 v0.11.3 & v0.11.4 are here! 🚀

🎉 AG2 Beta — ground-up redesign for production agents: streaming, multi-provider LLM support, typed tools with DI, middleware pipeline, first-class testing
🐟 TinyFish agentic web scraper
🔍 QuickResearchTool for parallel web research
🛡️ Security hardening: shell injection, path traversal, credential redaction
🔄 Mistral V2 + Gemini streaming

🙌 Thanks to our contributors!

Release: https://t.co/Xef8UCdl2g

#AG2 #AI #MultiAgentSystems #AIAgents #OpenSource #Beta

[AG2 @ag2oss]

Most teams don't lack AI. They lack working AI workflows.

McKinsey's 2025 State of AI: 80% of orgs use AI somewhere, but only ~20% redesigned workflows where value is created. Ideas aren't the bottleneck. Durable execution is.

Studio inside AgentOS closes that gap. Describe your workflow naturally. CaptainAgent builds the first agent graph. Test it live, iterate visually on canvas (triggers, agents, conditions, human review), then export AG2 code.

One workflow that "should already be automated"? Build + test it now: https://t.co/n6jsGiedJK

#ag2 #agentos #workflows #captainagent

[AG2 @ag2oss]

Introducing AG2 Beta - a new development track inside AG2.

A ground-up redesign of the agent experience with a production-focused feature set. Simplified API, streaming-first, async-first, event-driven, with dependency injection, and middleware.

Developed in parallel, so the current AG2 framework remains fully maintained and supported. Beta agents already work inside existing AG2 group chats and workflows.

5 LLM providers. Powerful new multi-agent orchestrations, observability, MCP, A2A, and AG-UI coming next.

We're just getting started, find out more here:
https://t.co/DgH0pJNkjM

#ag2 #oss #buildinpublic #opensource

[AG2 @ag2oss]

Company due diligence is still largely manual.

Analysts jump between websites, founder profiles, investor pages, and press coverage to piece together a company story.

AG2 enables structured multi-agent research workflows instead of single prompts.

Agents can also invoke tools when needed — like TinyFish for crawling sites and extracting data.

Full breakdown:
https://t.co/BKyxjT6a82