Clarion Street

The Open-Source Agent Stack Is Fragmenting — and That's the Point

·7 min read

For two years the “agentic orchestration” story was a three-way fight between the batteries-included frameworks: LangChain, AutoGen, and CrewAI. Each promised to turn a model into an agent with a few lines of glue. That framing is stale now. As of late August 2026, two of those three incumbents have either stalled or been absorbed, and the energy has moved somewhere smaller and harder to brand. The question worth asking is not who wins the framework war. It is whether the thing we have been calling “the infrastructure” is actually three different things, and which of them still matters.

The old battle ended quietly

AutoGen, at 60.7K stars, is in maintenance mode. Microsoft’s README says it “will not receive new features or enhancements and is community managed going forward,” and points new users to Microsoft Agent Framework as the “enterprise-ready successor.” AutoGen is licensed CC-BY-4.0, which is not an OSI-approved software license, and that detail matters because it is part of why a fork exists. AG2, at 4.9K stars, is that fork: “formerly AutoGen,” rebranded as “the Open-Source AgentOS,” now Apache-2.0 and independent of Microsoft.

OpenAI’s Swarm, at 21.9K stars, has the same footnote. Its README calls the project “experimental, educational” and says it “is now replaced by the OpenAI Agents SDK, which is a production-ready evolution of Swarm.”

CrewAI, at 57.7K stars, is still active and MIT-licensed. LangChain, at 145K stars, has repositioned itself as “the agent engineering platform,” with LangGraph, at 40.6K stars, as its low-level orchestration engine.

The incumbents did not lose to each other. AutoGen lost to Microsoft’s own successor. Swarm lost to OpenAI’s own SDK. LangChain survived by turning into a platform with LangGraph as the loop underneath.

Every provider ships its own SDK now

Within about seven weeks in early 2025, three model providers shipped thin, code-first agent SDKs: OpenAI Agents SDK (March 2025, MIT, 29K stars), Google ADK (April 2025, Apache-2.0, 21.3K stars), and Microsoft Agent Framework (April 2025, MIT, 13.2K stars, in .NET, Python, and Go). OpenAI’s Agents SDK says outright that it is “provider-agnostic, supporting the OpenAI Responses and Chat Completions APIs, as well as 100+ other LLMs.” The vendor SDKs are competing as neutral frameworks, not walled gardens. The same companies that sell the models are giving away the orchestration layer and telling you it works with everyone else’s model too.

The small and the typed

Next to the SDKs, a modular and minimalist camp is growing. Hugging Face’s smolagents (29K stars) sells “agents that think in code” and says the core logic fits in roughly a thousand lines. Pydantic AI (19.5K stars, MIT) is the typed-Python option. On the TypeScript side there is Mastra (27.5K stars) and the Vercel AI SDK (26.5K stars). Memory has split off into its own layer entirely, led by Mem0 (64.2K stars, “universal memory layer for AI Agents”) and Letta (24.5K stars, “stateful agents” with self-improving memory), with Zep behind them.

The interfaces became protocols

The tool and agent interfaces standardized while the frameworks churned. Anthropic open-sourced the Model Context Protocol in November 2024 as “a new standard for connecting AI assistants to the systems where data lives.” Google launched Agent2Agent in April 2025 with more than 50 partners, explicitly to “complement Anthropic’s Model Context Protocol,” for agent-to-agent coordination. A2A now sits with the Agentic AI Foundation, a Linux Foundation body, and we covered that governance move separately when it happened. The point for this piece is narrower. The tool-calling interface, one of the three candidates for “the infrastructure,” has already left the framework layer entirely and become a spec.

All three, separating

Is the infrastructure the model, the orchestration loop, or the tool-calling interface? OpenAI’s own launch split it three ways: the Responses API (the model-and-tool surface), built-in tools (the tool-calling interface), and the Agents SDK (the orchestration loop). It is all three, and they are separating.

The loop is where durability lives. LangGraph advertises “durable execution,” resuming from exactly where an agent left off. Agent Framework advertises “production-grade multi-agent workflows.” The framework layer is the least defensible of the three. Anyone can write one in a thousand lines, which is why the providers give it away for free.

Fragmentation is the strategy

Here is the thesis, and I want to flag that it is my read of the facts rather than a sourced claim. The providers did not fragment the framework layer by accident. They pushed it toward commodity on purpose, free MIT and Apache SDKs, while standardizing the interfaces they care about, MCP for tools and A2A for agents, under neutral foundations. Fragmentation of the framework layer is a feature. It stops any single framework from owning the abstraction layer the way LangChain briefly did in 2023 and 2024.

The cost is real. There is an integration tax, and there is “which of fifteen libraries do I pick” paralysis. That paralysis is the opening the niche libraries are filling.

What still has to be earned

If the framework is a commodity, the surviving differentiation is in the loop and the memory, because those encode state and reliability rather than prompt plumbing. The harness thesis, which we wrote about separately, restates cleanly here: the model is the commodity, and the loop and the memory are the product. LangGraph’s durable execution and Agent Framework’s workflow engine are not things you rewrite in a weekend, and memory is where an agent’s behavior stops being stateless.

The caveats

Star counts are not adoption. Everything above is GitHub stars pulled on August 28, 2026, and they measure attention, not production installs. Do not read them as market share.

“Fragmentation” is partly a rebrand. LangChain still has the largest ecosystem, and LangGraph is its orchestration core, so the big-framework-versus-small-library split is false at the edges. LangGraph is the low-level loop the niche libraries get compared to.

Claude Code is not open source. It is the most-starred agentic tool at 143K stars, and it ships with no license at all, source-available rather than open source. The single most popular agentic tool sits outside the open-source battle this post is about, which should cut against any clean “open source wins” ending.

Some of these claims rest on a single vendor-authored source, a README or API metadata. AG2’s branding, smolagents’ line count, Mem0 and Letta’s descriptions, and the A2A governance phrasing are self-description. Read them as marketing rather than independent assessment.

The long tail keeps growing. Agno (42K stars, formerly Phidata), LlamaIndex’s pivot toward a “document agent and OCR platform,” and OpenHands (85K stars, MIT), the largest open agentic-coding framework. This is not a five-horse race.

What I would tell someone picking a stack

Pick a provider SDK or a small code-first library for the orchestration loop. Adopt MCP and A2A for the interfaces. Treat the batteries-included framework as optional glue, not the foundation. The thing that was the whole story two years ago is now the least interesting part of the stack.

Bottom line

The open-source agent stack is fragmenting, and it is not a sign of failure. The providers are making the framework layer cheap on purpose so that no single project owns it, while the real competition moves to the orchestration loop, the memory layer, and the two protocols. The frameworks everyone argued about in 2024 have either been absorbed by their own vendor’s SDK or become platforms. The layers that still matter are the ones you cannot trivially rewrite: state, reliability, and the interfaces everything else plugs into.

Sources

  1. langchain-ai/langchain — GitHub
  2. microsoft/autogen — GitHub
  3. crewAIInc/crewAI — GitHub
  4. openai/swarm — GitHub
  5. openai/openai-agents-python — GitHub
  6. microsoft/agent-framework — GitHub
  7. google/adk-python — GitHub
  8. ag2ai/ag2 — GitHub
  9. langchain-ai/langgraph — GitHub
  10. huggingface/smolagents — GitHub
  11. pydantic/pydantic-ai — GitHub
  12. mastra-ai/mastra — GitHub
  13. vercel/ai — GitHub
  14. mem0ai/mem0 — GitHub
  15. letta-ai/letta — GitHub
  16. getzep/zep — GitHub
  17. anthropics/claude-code — GitHub
  18. All-Hands-AI/OpenHands — GitHub
  19. agno-agi/agno — GitHub
  20. run-llama/llama_index — GitHub
  21. New tools for building agents — OpenAI, March 11, 2025
  22. Introducing the Model Context Protocol — Anthropic, November 25, 2024
  23. A2A: A new era of agent interoperability — Google Developers Blog, April 9, 2025
  24. A2A Protocol