AG2 1.3.0 adds delegation-tree observers and requires Python 3.11

Updated Nov 4, 2026AG2

Summary

AG2 1.3.0 (November 4, 2026) lets coordinators observe a whole delegation tree, adds scoped variables and a shared task lifecycle, and raises the Python floor to 3.11 before 1.4 removals.

Tool: AG2
Category: Frameworks
Date: Nov 4, 2026
Open source: Yes
License: Apache-2.0
Pricing: Not stated in the sources we cited.
Sources: GitHub release v1.3.0
GitHub repository
AG2 docs
PyPI: ag2

AG2 1.3.0 was released on GitHub on November 4, 2026 and uploaded to PyPI the same day (09:22 UTC). It is a feature release with a hard environment change: Python 3.11 or newer is required. Several 1.0-era entry points still work here with deprecation warnings and are scheduled for removal in 1.4.0.

Release facts

Fact	Detail
Product	AG2, formerly AutoGen
Version	1.3.0
Release date	November 4, 2026 (GitHub release; PyPI upload 09:22 UTC)
Previous release	1.2.2, October 27, 2026
License	Apache-2.0, as reported by the GitHub API
Python	`>=3.11`, per the 1.3.0 notes and the PyPI `requires_python` field
Install	`pip install -U "ag2[openai]==1.3.0"` (swap the extra for your provider)

What changed

Observe the delegation tree. Ancestors receive events from every descendant, including streamed chunks, as ChildEvent observations. Those events do not enter the ancestor's model transcript or trigger its tools, model loop or human-in-the-loop handlers. Typed @observer(...) callbacks can monitor the tree, and await event.origin.send(...) delivers a reply through the child invocation's normal handlers. Handles expire when that invocation ends.
Explicit variable scopes. ScopedValue controls how state crosses a delegation boundary: COPIED (deep-copied per branch, the default for plain values), LOCAL (not inherited) and SHARED (one binding across the tree). Declare shared names at the root. The notes say applications provide their own synchronization for concurrent updates.
Dynamic context. statusbar_middleware presents current facts such as a plan, progress or remaining attempts before the next model call. persist=True keeps snapshots as service observations.
Shared task lifecycle. Model runs, delegations and standalone Task blocks share one lifecycle. Descendant usage arrives as original UsageEvent records rather than rolled-up kind="subtask" totals.
History queries. Saved events can be filtered by type and condition, including since_last_user_input(). Queries read retained top-level events; records removed by compaction are gone.

Deprecations to fix before 1.4

These APIs still run in 1.3 with warnings and are listed for removal in 1.4.0:

Deprecated API	Replacement named in the notes
`TaskConfig`, `Agent(tasks=...)`, `Agent.tasks`, and the injected `run_subtask` / `run_subtasks` tools	Build workers explicitly and delegate with `Agent.as_tool()`
`Agent(assembly=...)`	Register `AssemblerMiddleware` through `middleware=`; add `HaltCheckMiddleware` when using `AlertPolicy`
`Agent(knowledge=...)`	Use `plugins=[KnowledgePlugin(...)]`
`Agent.task()`	Construct `Task(owner_id=..., spec=TaskSpec(...))` directly

KnowledgeConfig, assembly policies and Task remain supported; the deprecations apply to the entry points above. As previously announced, OpenAIConfig becomes an alias for OpenAIResponsesConfig in 1.4. The notes recommend OpenAIResponsesConfig or OpenAICompletionsConfig explicitly.

Who should upgrade now

Upgrade the Python runtime first if any environment is still on 3.10: 1.3.0 will not install there. Teams that coordinate several workers get the observer and scoped-variable APIs in this release. If your code still constructs agents with tasks=, assembly= or knowledge=, migrate those call sites on 1.3 while they only warn, rather than waiting for the 1.4 removals.

Alternatives

AG2 compared with its alternatives
Tool	License	Pricing note
AG2This launchAG2 (formerly AutoGen) is an Apache-2.0 Python framework for AI agents and multi-agent networks, installed from PyPI as ag2 and imported as ag2.	Apache-2.0	AG2 is Apache-2.0 and free to install from PyPI. You pay your model provider for tokens.
CrewAICrewAI is a framework for orchestrating role-playing, autonomous AI agents that work together.	MIT	Not stated
LangGraphLangGraph is an open-source orchestration runtime for building stateful, resilient agents.	MIT	Not stated
OpenAI AgentsOpenAI's agent harness, shipped as a Python SDK update in April 2026 and as the hosted Agents API in September 2026.	Open source	SDK and Agents API use standard token and tool pricing. The Agents API has no additional fee.
AgnoAgno is an orchestration and service layer for agents, exposing them through a REST API and an MCP server.	Apache-2.0	Not stated

Questions

What is AG2?

AG2 is the open-source agent framework that grew out of AutoGen. Its README calls it an open-source AgentOS for building AI agents and getting several agents to cooperate on tasks. It is maintained by volunteers from several organizations.

Is AG2 the same as AutoGen?

It started as AutoGen. Since AG2 v1.0 the package imports as ag2 with an Agent class and Network for multi-agent work. The older autogen namespace, with ConversableAgent and GroupChat, moved to AG2 Classic (ag2ai/ag2-classic, pip install ag2-classic), which is still maintained.

How do I install AG2?

pip install ag2 with the extra for your provider, such as ag2[openai], ag2[anthropic], ag2[gemini] or ag2[ollama]. AG2 1.3.0 requires Python 3.11 or newer, matching the release notes and the package's requires_python field on PyPI.

Is AG2 open source?

Yes. The GitHub API reports Apache-2.0 for ag2ai/ag2.

What is the latest launch covered here?

AG2 1.3.0, published November 4, 2026. It requires Python 3.11, adds observers for a whole delegation tree and scoped variables, and deprecates several 1.0 entry points ahead of removal in 1.4.0.

Filed on Agent Radar. Claims in this note are limited to the sources above.