AG2

AG2 v1.0 is built around an async Agent that calls a model provider, runs the tool-calling loop and publishes every step (model calls, tool calls, human-input requests) as events on a stream. Optional harness components add context assembly policies, a path-based KnowledgeStore for memory across runs, history compaction and sub-task delegation. Middleware covers retries, token budgets and similar concerns.

Multi-agent work has two routes. Within one agent, other agents can be attached as tools with Agent.as_tool(); the calling model decides when to delegate. For coordinated teams, ag2.network sets up a hub-and-spoke network: the hub holds the registry, per-channel write-ahead logs, governance rules and an audit log, and agents exchange envelopes over channels whose adapter sets the rules (free two-party conversation, one-question consulting, round-robin discussion, or a workflow driven by a declarative, JSON-serialisable TransitionGraph). The workflow adapter is the documented replacement for the classic GroupChat. The hub runs in-process by default or over WebSocket for multi-process deployments.

State is externalised behind History, Storage and Stream protocols that can be backed by Redis or a database. agent.resume() rebuilds a turn from recorded events, and network channel transcripts can be replayed from the write-ahead log.

A tool can call context.input(...) to pause the run and ask a person; the agent's hitl_hook decides how the question is answered (CLI prompt, web UI, queue). The built-in approval_required() tool middleware asks the user to approve or deny a specific tool call before it runs, with an option to always allow that tool for the rest of the conversation. In a network, a HumanClient joins channels as a non-LLM participant that your UI drives. Over AG-UI, a run can end with an interrupt and be resumed by a later run on the same thread, and a gated tool call can be sent to the frontend for approval. When AG2 drives a CLI coding agent over ACP, its permission requests go to the same hook unless permission_policy is set to auto-approve or deny.

pip install "ag2[openai]"

import asyncio
from ag2 import Agent
from ag2.config import OpenAIConfig

config = OpenAIConfig("gpt-4o-mini")  # reads OPENAI_API_KEY
researcher = Agent("researcher", prompt="Return three short factual findings.", config=config)
writer = Agent("writer", prompt="Turn the notes you get into one clear paragraph.", config=config)
coordinator = Agent("coordinator", config=config,
    prompt="Delegate research first, then pass the findings to the writer.",
    tools=[researcher.as_tool(description="Research a topic and return findings."),
           writer.as_tool(description="Write prose from notes passed as context.")])

async def main() -> None:
    reply = await coordinator.ask("Write a short note on the A2A protocol.")
    print(reply.body)

asyncio.run(main())

• Requires Python >= 3.10. Only minimal dependencies install by default; add the extra for your provider (ag2[openai], ag2[anthropic], ag2[gemini], ...).
• Provider configs read the standard environment variable (OPENAI_API_KEY, ANTHROPIC_API_KEY, GEMINI_API_KEY).
• pip install ag2 (v1.0+) no longer ships the autogen import name or ConversableAgent / GroupChat; code using them needs pip install ag2-classic, and moving it to v1.0 is a rewrite, not an upgrade (see the group chat migration guide).
• The API is async throughout; call ask() from inside an event loop.
• Protocol features are extras: ag2[mcp], ag2[a2a] (serving also needs a2a-sdk[http-server] or a2a-sdk[grpc]), ag2[ag-ui], ag2[acp].

AG2's docs say it diverged from the AutoGen codebase in November 2024. In v1.0 it moved the AutoGen-style API to ag2-classic and switched to an async Agent and a hub-based network, while Microsoft's AutoGen is in maintenance mode.

Yes, through AG2 Classic: pip install ag2-classic, documented at classic.docs.ag2.ai. The ag2 package from v1.0 no longer includes that namespace.

Yes, all three are first-party. MCP works as client and server, A2A as client and server, and AGUIStream exposes agents to AG-UI frontends. Each needs its own install extra.

Yes. With ag2[acp], ClaudeCodeConfig, CodexConfig and OpenCodeConfig launch those CLIs as ACP agents, and AG2 streams their messages, tool calls and permission prompts.

The framework is Apache-2.0. The AG2 website also advertises hosted products behind a request-access form without public pricing, so the pricing model is listed as unknown.