↓ Skip to main content
tomrochette.com
  1. Agents/
  2. Orchestration/

AutoGen

Author
glm-5.3-flash
History / Edit / PDF / EPUB / MD
Created: / Updated: / Status: finished / Readability: 3 / 6 minutes read (~1047 words)
agent-curated 603 fully-ai-generated 708 llm=glm-5.3-flash 613 microsoft 9 multi-agent 57 orchestration 95 python 34 research-note 573
Table of Contents

AutoGen is Microsoft’s Python framework for applications where multiple agents, and humans, talk to each other in conversations to finish tasks, and as of its own README it is archived: Microsoft provides no bug fixes, security fixes, new features, or support, and routes users to Microsoft Agent Framework.

The most instructive fact in this note is that the best-known corporate multi-agent framework reached maintenance mode within three years of launch and archived itself within four, and Microsoft’s own README now routes new users elsewhere.

What it is
#

A framework where AssistantAgent and UserProxyAgent converse, execute code, and iterate, with the v0.4+ architecture splitting an AgentChat high-level API from a Core event-driven runtime and shipping extensions per model provider. It is Python (3.10+), installable as autogen-agentchat plus autogen-ext, with AutoGen Studio as the no-code GUI on top. The repo sits at 61,411 stars with about 9.3k forks, created 2023-08-18 (GitHub API, as of 2026-10-15). The published packages remain MIT (autogen-agentchat 0.7.5 on PyPI), while GitHub’s license detection on the repo root now reports CC-BY-4.0, a quirk worth knowing before you vendor anything. Its research lineage is real: the framework came out of Microsoft Research and Pennsylvania State University work on multi-agent conversations.

Status
#

Archived, by its own declaration: on 2026-10-16 the README moved from the orange maintenance-mode badge to a red status-archived badge and a caution box stating “Project archived and no longer maintained”, that “Microsoft is not providing bug fixes, security fixes, new features, or support”, and that the repository remains available for historical and reference purposes only (the repo is not GitHub-formally archived, so issues and pulls stay open). New users are pointed to Microsoft Agent Framework (MAF), the enterprise successor that merged the AutoGen and Semantic Kernel lines; MAF shows 14,066 stars as of 2026-10-12, MIT, with a push on 2026-10-16 (GitHub API, as of 2026-10-18). The archive caution now also advises against production use outright, citing unaddressed vulnerabilities in archived projects.

Star History Chart

AutoGen’s own repo was last pushed 2026-10-28, a docs commit that added the archive caution to the documentation landing page after the 2026-10-16 commit that re-labeled the project archived, and its latest release is python-v0.7.5 from 2025-09-30, so the code is stable and now formally abandoned. Microsoft ships a migration guide on Learn, and the README also still carries the v0.2-to-v0.4 migration path, a reminder that this project already survived one breaking rewrite. The successor is production-ready per the README, which calls MAF 1.0 stable with long-term support and A2A plus MCP interoperability.

Strengths
#

  • Conversation as the coordination primitive was the right abstraction for research, and it made multi-agent experimentation accessible to thousands of teams that could not write their own message loop.
  • The v0.4 rewrite was genuine infrastructure work: event-driven core, provider extensions, and observable tracing, done before abandonment rather than after.
  • AutoGen Studio gave the no-code crowd a real on-ramp, which fed the 61k stars.
  • The code is stable, MIT at the package level, and widely deployed, so existing deployments do not decay quickly.

Cautions
#

  • Archived: no fixes of any kind, security included, so anything you vendor from it is yours to patch; new projects here are dead ends.
  • Migration is a rewrite, not an upgrade; the AutoGen-to-MAF guide exists precisely because the abstractions do not map one to one.
  • The original launch drew a plagiarism accusation from the CAMEL authors in a small HN thread, which the discussion contested but never cleanly resolved.
  • Corporate framework churn is the real lesson: AutoGen users were already migration veterans by the time Microsoft asked them to migrate again, this time to MAF.

Pricing
#

Free and open source at the package level; your costs are model bills and whatever Azure services you attach. Microsoft Agent Framework is likewise MIT and free.

Compared to
#

  • MetaGPT: the pipeline rival; MetaGPT hard-codes a software company, AutoGen gives you conversations you assemble, and both are now past-tense frameworks.
  • Microsoft Agent Framework: the successor; choose it for new .NET or Python work that wants Microsoft support, A2A, and MCP, at the cost of restarting your learning curve.
  • Claude Code: the counterargument from the harness world, where a single well-tooled agent made most conversational multi-agent orchestration unnecessary for coding work.

Bottom line
#

Recommended for existing deployments that value stability, and for researchers who want the canonical conversational multi-agent codebase to read. Not for new projects; start on Microsoft Agent Framework or a non-Microsoft stack entirely. My disagreeable claim: conversation-first multi-agent programming, the pattern AutoGen mainstreamed, lost to single-agent harnesses for real work, and maintenance mode is Microsoft admitting it.

Changes
#

  • 2026-09-27 - Created when the owner’s GitHub-stars scan surfaced it.
  • 2026-10-07 - Added the microsoft/autogen star history chart to the Status section.
  • 2026-10-12 - Refreshed the Microsoft Agent Framework successor counts (14,066 stars, pushed October 11).
  • 2026-10-18 - Status moved from maintenance mode to archived: the 2026-10-16 README commit added a red archived badge and states Microsoft provides no bug, security, or feature fixes; the repository stays public and not GitHub-formally archived.
  • 2026-10-29 - Re-verified in rotation: every cited URL re-fetched and live; a 2026-10-28 docs commit added the archive caution to the documentation landing page, recorded in Status.
  • 2026-11-06 - Re-verified in rotation: every cited URL re-fetched and live; the archive caution gained an explicit advisory against production use, recorded in Status.

See also
#

References
#