Short answer: the biggest change in AI agent frameworks this fall is that the most-starred project in the category is barely two months old — and it isn't a framework. DeepSeek Harness, open-sourced on August 13, now has ~248,000 GitHub stars (verified via the GitHub API on Oct 12), more than LangChain's ~148,000. It's pitched as a harness — a runtime where the model, tools, skills and UI are all plugins — not a framework you write code inside. That reframing is the story. Here's the state of play, the releases that shipped in the last two weeks, and what to actually build on.
Here's the whole thing in one screen:
- A newcomer took the crown — on attention, not production use. DeepSeek Harness is an MIT-licensed developer preview. Stars measure buzz; LangGraph, CrewAI and the OpenAI Agents SDK are still what teams ship on.
- "Harness" is the new word. Strands renamed its SDK to
harness-sdk; Pydantic shipped apydantic-ai-harness. The idea: the runtime sits above the model. - Microsoft's consolidation is done. AutoGen is frozen; the unified Agent Framework (Python + .NET) is the live successor.
- Everyone else is iterating weekly — CrewAI, Agno, Pydantic AI, LangGraph and Google ADK all shipped releases in the last ~10 days. The quiet one is Hugging Face's smolagents (no release since May 29).
The disruptor: DeepSeek Harness#
DeepSeek Harness (the CLI is dsh) is a Node.js/TypeScript agent runtime built on a meta-framework called Cordis, where "everything is a plugin": model adapters, tool registries, sandboxes, session state, event dispatchers and even the UI all load as independent extensions. InfoQ framed it as the start of modular, unbundled agent infrastructure — the runtime decoupled from any one model. It already has a sprawling ecosystem of community plugin hubs and a desktop build.
The honest read: ~248k stars in two months is a genuine phenomenon, and also a reason for discipline. It's a developer preview; APIs will change, and a star is a bookmark, not a production deployment. Prototype with it if the plugin model appeals. Don't make it your default the way you would a 1.x framework with a year of battle-testing behind it.
The real trend: "framework" is becoming "harness"#
A framework is something your agent lives inside — you code against its abstractions. A harness inverts that: it's a runtime above the model, orchestrating swappable plugins. DeepSeek Harness is the loudest example, but the language is spreading. AWS's Strands renamed its SDK to harness-sdk ("build an agent harness and control it end-to-end… any model, any cloud"), and Pydantic now ships a pydantic-ai-harness alongside its main library. The appeal is portability — change the model or the tools without rewriting the agent — which is the same instinct behind keeping your inference layer swappable: own the boundary, don't marry one model. The tradeoff is maturity; the pattern is new.
The incumbents, as of mid-October#
LangChain / LangGraph remain the center of gravity. LangChain (~148k stars) is now billed as "the agent engineering platform"; the production workhorse is LangGraph (~43k), whose stateful, graph-based orchestration shipped 1.2.14 on Oct 6 on a near-weekly cadence. If you need to debug a controlled, multi-step agent in production, this is still the default.
CrewAI (~60k) shipped 1.15.27 on Oct 9 and is the pick for role-based "crews" of cooperating agents. Agno (~43k, 3.1.2 on Oct 8) repositioned from a library to a "platform" for high-performance Python agents. Pydantic AI (~21k, 2.55.0 on Oct 9) is the type-safe choice — validated, typed I/O end to end. Google ADK (~22k, 2.11.0 on Oct 2) is the code-first pick for Gemini and Vertex.
The OpenAI Agents SDK (~30k) is the lightweight option — handoffs, guardrails, tracing — and shipped 0.23.1 on Oct 2, but note it's still pre-1.0 despite heavy adoption. Microsoft finished its consolidation: AutoGen (~61k) hasn't taken a feature commit since April, and development moved to the unified Microsoft Agent Framework (Python + .NET; Python 1.21.0 on Oct 8). Starting fresh on the Microsoft stack means starting there, not on AutoGen. And LlamaIndex (~52k) repositioned toward being "the document processing platform for AI" — data and RAG agents rather than general orchestration.
What to build on#
Match the tool to the job, not the leaderboard: LangGraph for production control, CrewAI for role-based crews, Pydantic AI for type safety, the OpenAI Agents SDK for a lightweight SDK, Microsoft Agent Framework for .NET/Azure, Google ADK for Gemini. Prototype DeepSeek Harness to feel out the plugin model — just not as your production default yet. For the longer view of how these ranked before this fall's churn, see the frameworks ranked by stars in August; the top of that list has already been rewritten.
