Summary
The most starred AI/ML repos (50k+ stars) are dominated by foundational infrastructure tools (Ollama, LangChain, Hugging Face Transformers) and emerging agent platforms (OpenClaw, Dify, Hermes Agent). AI agent frameworks are the fastest-growing category. This report covers the top repos by stars, rising frameworks, established agent toolkits, and the key trends shaping the space.
1. Most Starred AI/ML Repositories (50k+ Stars)
The table below lists the most starred AI/ML repositories on GitHub as of September 2026. Stars reflect both longevity and community adoption.
| Repository | Approx. Stars | Category |
|---|---|---|
| OpenClaw | ~366k | Agent Platform |
| ECC | ~268k | Agent Framework |
| DeepSeek Harness | ~236k | Agent Evaluation |
| Firecrawl | ~185k | Web Scraping / Agent |
| Ollama | ~160k+ | LLM Runtime |
| Dify | ~155k | Agent Platform |
| Ponytail | ~146k | Agent Framework |
| LangChain | ~147k | Agent Framework |
| Hermes Agent | ~141k | Agent Platform |
| Langflow | ~140k+ | Agent Flow Builder |
| Hugging Face Transformers | ~130k+ | ML Library |
| LLMs-from-scratch | ~102k | Education |
| DeepSeek-V3 | ~100k+ | LLM Model |
| Browser Use | ~91k | Browser Automation Agent |
| RAGFlow | ~91k | RAG Framework |
| llama.cpp | ~80k+ | LLM Inference |
| OpenHands | ~87k | Coding Agent |
| Unsloth | ~76k | LLM Fine-tuning |
| MetaGPT | ~70k | Multi-agent Framework |
| Mem0 | ~66k | Agent Memory |
| GitHub Spec Kit | ~55k | Coding Agent |
What stands out: OpenClaw leads the pack at ~366k stars, a position it has gained rapidly as a self-improving agent platform. The list is split between foundational infrastructure (Ollama, Transformers, llama.cpp) and the newer wave of agent platforms (OpenClaw, Dify, Hermes Agent). Several repos in the 50k–100k range are focused on specific agent capabilities — memory (Mem0), browser automation (Browser Use), and coding (OpenHands, GitHub Spec Kit).
2. New and Trending AI Agent Frameworks
These frameworks have gained significant traction in 2025–2026 and represent the rising edge of the agent ecosystem.
| Framework | Focus Area |
|---|---|
| OpenClaw | Self-improving agent platform |
| Hermes Agent | General-purpose agent platform |
| GSD (link unverifiable — no matching public repo found) | Goal-directed agent execution |
| ECC | Event-driven agent coordination |
| DeepSeek Harness | Agent evaluation and benchmarking |
| Ponytail | Agent orchestration |
| Browser Use | Browser-based autonomous agents |
| Orca | Multi-modal agent reasoning |
| Firecrawl | Web scraping and agent-driven data extraction |
| Mastra | TypeScript-first agent framework |
| DeerFlow | Research and knowledge-agent workflows |
| smolagents | Lightweight, minimal agent framework |
| Letta | Agent memory and state management |
| Unbody | Agent-native database and memory layer |
| OWL | Modular agent composition |
What stands out: TypeScript-first frameworks like Mastra are gaining ground, reflecting a broader trend of the agent ecosystem moving toward type-safe, developer-friendly tooling. Several of these frameworks (OpenClaw, Hermes Agent, ECC) emphasize self-improvement and autonomous capability — a shift from simple orchestration to agents that can evaluate and refine their own behavior.
3. Established Agent Frameworks and Toolkits
3.1 General-Purpose Agent Frameworks
| Framework | Description |
|---|---|
| CrewAI | Multi-agent orchestration with role-based task delegation |
| AutoGen | Microsoft's framework for multi-agent conversations and collaboration |
| LangGraph | Stateful, graph-based agent workflow builder (from LangChain team) |
| OpenAI Agents SDK | Official OpenAI SDK for building agentic workflows |
| Google ADK | Google's Agent Development Kit for building AI agents |
| PydanticAI | Agent framework built on Pydantic for structured data and validation |
| Agno | Multi-agent framework with built-in knowledge and tools |
| Semantic Kernel | Microsoft's SDK for integrating AI agents with enterprise apps |
| LlamaIndex | Data framework for building RAG-powered agents |
| Haystack | End-to-end NLP and agent pipeline framework |
| Rig | Minimalist agent framework focused on composability |
| OpenAI Swarm | Lightweight framework for agent handoffs and routing |
| BabyAGI | Simple autonomous agent loop for task generation and execution |
3.2 AI Coding Agents
| Tool | Description |
|---|---|
| Claude Code | Anthropic's agentic coding assistant |
| GitHub Spec Kit | GitHub's official toolkit for spec-driven agent development |
| OpenHands | Open-source coding agent capable of executing code in sandboxed environments |
| Cline | AI coding assistant with terminal and file-editing capabilities |
| Continue | Open-source AI coding assistant for VS Code and JetBrains |
| Cursor | AI-first code editor with integrated agent capabilities |
| Devin (proprietary — Cognition Labs; no public repo) | Autonomous AI software engineer |
3.3 Autonomous Agents
| Agent | Description |
|---|---|
| AutoGPT | Pioneering autonomous agent that breaks tasks into subtasks and executes them |
| AgentGPT | Web-based autonomous agent that plans and executes goals |
| Huginn | Self-hosted automation platform for building agent-like workflows |
| n8n | Workflow automation tool with AI agent nodes |
| RAGFlow | Agent-powered RAG system for document understanding |
| Pathway | Stream processing framework adapted for agent data pipelines |
4. Key Trends
4.1 Self-Improvement Is the New Frontier
The most starred and fastest-growing repos (OpenClaw, Hermes Agent, ECC) share a common theme: agents that can evaluate, critique, and improve their own outputs. This represents a shift from static orchestration frameworks to dynamic systems where agents iterate on their own performance. The DeepSeek Harness repo (~236k stars) exists specifically to benchmark and evaluate this capability.
4.2 Security Is Lagging Behind Adoption
As agent frameworks proliferate, the security tooling and best practices have not kept pace. Most agent repos lack built-in sandboxing, permission controls, or audit logging. The gap between what agents can do (execute code, browse the web, modify files) and how safely they are contained is a growing concern for production deployments.
4.3 TypeScript Is Having Its Moment
Frameworks like Mastra and several newer entrants are TypeScript-first, reflecting the broader JS/TS ecosystem's influence on AI tooling. This contrasts with the Python-dominated ML landscape and signals that agent development is becoming more of a software engineering discipline than a data science one.
4.4 Specialization Wins
General-purpose frameworks (LangChain, AutoGen) still dominate in stars, but the fastest-growing repos are specialized: browser automation (Browser Use), memory (Mem0), evaluation (DeepSeek Harness), and coding (OpenHands). The trend is toward narrow, well-defined agent capabilities rather than monolithic frameworks.
4.5 MCP Is the New Integration Standard
The Model Context Protocol (MCP) is emerging as the standard way for agents to connect to external tools, data sources, and services. This mirrors how REST APIs became the standard for web integration in the 2010s. Repos that support MCP natively tend to attract more adoption and community contributions.
4.6 Two Lanes: Personal AI Assistants vs Developer Frameworks
The ecosystem has split into two distinct lanes:
- Personal AI assistants — tools like Cursor, Cline, and Continue that augment individual developers with AI-powered coding help.
- Developer frameworks — tools like LangChain, CrewAI, and AutoGen that provide the building blocks for constructing multi-agent systems and complex workflows.
Both lanes are growing, but they serve different audiences and have different adoption patterns.
5. Limitations and Open Questions
- Star counts are approximate and can fluctuate significantly day to day. The figures here are based on research conducted in the week of 2026-09-27 and should be treated as indicative rather than exact.
- The list of "trending" frameworks is not exhaustive. Many smaller, niche agent repos may be growing quickly but do not appear in the top-starred lists.
- The distinction between "agent frameworks" and "general ML libraries" is blurry. Some repos (e.g., Hugging Face Transformers) are foundational infrastructure that agents build on, but are not agent-specific.
- Security trends are based on general observations from the ecosystem, not a systematic audit of individual repos.
Sources
- GitHub repository star counts and metadata, accessed week of 2026-09-27
- Community observations on agent framework adoption and trends in 2025–2026
- Repository descriptions and documentation from individual project pages