Top 10 GitHub Repositories for Building AI Agents in 2026

Short version: since our April list, the centre of gravity moved. The question is no longer "which framework", it is "which runtime, which harness, which skills". Several repos that did not exist a year ago now sit above 100,000 stars, two April picks are effectively unmaintained, and one (Flowise) was archived in August. Every number below was pulled from the GitHub API on 31 October 2026, not copied from another list, and this page is refreshed monthly with a visible changelog.

1. openclaw/openclaw (393,532 stars)

OpenClaw is the repo that redefined what "popular" means in this space. Created in November 2025, it crossed 390,000 stars in under a year, with 82,000 forks behind it, so this is not a star-farming artefact. The pitch is a personal AI agent that actually does things on your machine, any OS, any platform. It ships a release almost weekly (v2026.10.9 landed on 29 October) and carries close to 8,000 open issues, which tells you two things: the community is enormous, and the surface area moves fast. Treat it as a runtime you build on, not a library you pin and forget.

2. langchain-ai/langgraph (44,217 stars)

Our April post claimed 90,000 stars for LangGraph. The real figure is 44,217, and it is still the most production-proven way to build agents as explicit state graphs: checkpointing, durable execution, human-in-the-loop interrupts and time travel debugging come for free. It sits on the LangChain 1.x line (langchain-core 1.7.3 shipped on 29 October) and LangGraph itself is at 1.2.19. If you need an agent that can pause for a day, resume, and explain what it did, this is still the default answer.

3. crewAIInc/crewAI (60,338 stars)

CrewAI kept its role-based model (agents with goals and backstories working as a crew) and added Flows for event-driven control when a free-running crew is too loose. Release cadence is relentless: 1.15.33 on 29 October. Pick it when the problem decomposes naturally into roles, researcher, writer, reviewer, and you want readable code more than fine-grained control of every transition.

4. microsoft/agent-framework (14,413 stars)

This is the repo that replaced AutoGen on our list, and not by our choice: on 16 October 2026 Microsoft marked AutoGen archived and no longer maintained, and its README now advises against production use and points to Agent Framework and a migration guide. Agent Framework merges the ideas from AutoGen and Semantic Kernel into one framework with first-class Python and .NET support, graph-style workflows, multi-agent orchestration and A2A and MCP interoperability; python-1.24.0 shipped on 30 October 2026. The star count is modest because the repo is young, but if you are on Azure or in a .NET shop, this is where Microsoft's engineering effort is going.

5. openai/openai-agents-python (30,415 stars)

The Agents SDK is deliberately small: agents, tools, handoffs between agents, guardrails and built-in tracing. It works with non-OpenAI models through LiteLLM, so it is not the lock-in it looks like. v0.24.0 shipped on 15 October. The lightweight design is the point; teams that outgrow it usually move to LangGraph for durable state rather than fighting the SDK.

6. google/adk-python (22,059 stars)

Google's Agent Development Kit is code-first and opinionated about structure: hierarchies of agents, a tool ecosystem, an evaluation harness and a direct path to deployment on Vertex AI. v2.13.0 landed on 23 October. If Gemini is your primary model and you want eval and deployment solved on day one, ADK is the shortest route.

7. pydantic/pydantic-ai (20,968 stars)

Pydantic AI brings the Pydantic philosophy to agents: typed dependencies, validated structured outputs and model-agnostic providers, with a v2 line that now ships almost daily (v2.64.0 on 30 October). It is the framework we reach for when the agent's output has to be a schema, not prose, and when we want the type checker to catch mistakes before the LLM does.

8. browser-use/browser-use (119,645 stars)

A Python library that lets an agent see and operate a real browser, plus a hosted cloud with stealth browsers for teams that do not want to run browser fleets themselves. At 119,645 stars it is the most-starred capability library on this list, which reflects where agents create value right now: forms, dashboards and sites with no API. Release 0.13.14 shipped on 26 October.

9. OpenHands/OpenHands (92,338 stars)

OpenDevin became OpenHands and then moved to its own organisation, which is why the April link no longer resolves. It is the leading open-source coding-agent platform: sandboxed runtimes, multi-model support and an active 1.x line (v1.34.0 on 27 October). If you want to run a coding agent you control rather than rent one, start here.

10. modelcontextprotocol/servers (91,715 stars)

Not a framework, but every framework above speaks the Model Context Protocol, and this repo holds the reference servers: filesystem, git, fetch, memory and more. v2.2.0 shipped on 28 October. Learning the server pattern here is the quickest way to give any agent, in any framework, safe access to your own systems.

The skills and harness wave

The biggest shift since April is not a framework at all. The fastest-growing repos are skills, harnesses and fleet managers for coding agents like Claude Code and Codex. We keep them on a watch list rather than in the top ten because they are months old, but the fork counts show the adoption is real.

• morluto/rea (414,452 stars, 103,661 forks, created April 2026): "reverse engineer anything with agents", from app behaviour down to native binaries. The single most-starred repo in this entire survey.
• obra/superpowers (309,822 stars): an agentic skills framework and development methodology. Only three commits in October, so it looks settled rather than abandoned.
• affaan-m/ECC (293,783 stars, created January 2026): an agent harness performance system covering skills, instincts, memory and security.
• Graphify-Labs/graphify (139,843 stars): turns a codebase, its docs, SQL schemas and PDFs into a queryable knowledge graph for agents.
• addyosmani/agent-skills (112,122 stars): production-grade engineering skills for AI coding agents, releasing several times a week.
• stablyai/orca (111,529 stars): a workspace for running a fleet of agents side by side. Nearly 8,000 open issues, so expect rough edges.
• bytedance/deer-flow (84,760 stars): a long-horizon "super agent" harness that researches, codes and creates.

Low-code builders

If you would rather draw the agent than code it: n8n (209,197 stars, n8n@2.45.7) is the workflow platform with native AI nodes, Dify (160,086 stars, 1.18.0) builds agentic workflows and RAG pipelines, and Langflow (156,191 stars, v1.13.0) is the visual front end to the LangChain ecosystem. Flowise, which we recommended in the spring, was archived on GitHub in August 2026. If you are still on it, plan a migration; Langflow is the closest match. AutoGPT (187,844 stars) now lives on as the AutoGPT Platform, a visual agent builder; its classic Forge code has been marked unsupported since late 2024.

Dropped from the list, and why

• microsoft/autogen: still 61,646 stars, but the project was put into maintenance mode in April 2026 and formally archived on 16 October 2026, with Microsoft advising against production use. Our April figure of 160,000 stars was simply wrong, and so was calling it the most-starred agent framework.
• geekan/MetaGPT: last release v0.8.1 in April 2024, last push January 2026. Historically important, not where you should start today.
• microsoft/promptflow: last release promptflow_1.17.1 in January 2025. Still on GitHub, no longer something we would start a new project on.
• FlowiseAI/Flowise: archived August 2026.
• The April list also had CrewAI twice under two names and a niche Elixir port of LangChain in the top ten. Both were mistakes, now fixed.

How to choose

• You need durability and auditability: LangGraph.
• You think in teams of specialists: CrewAI.
• You are a .NET or Azure shop: Microsoft Agent Framework.
• You want the least code and good tracing: OpenAI Agents SDK.
• Gemini plus Vertex deployment: Google ADK.
• Outputs must match a schema: Pydantic AI.
• The job is in a browser or a codebase: browser-use or OpenHands.
• A personal always-on assistant: OpenClaw, with your eyes open about churn.
• Whatever you pick: expose your own systems through MCP servers so you are not locked to the framework.

How this list is built

Star counts, latest releases, creation dates and fork counts come straight from the GitHub API, captured on 31 October 2026. To make the top ten a repo must be about building agents, must have pushed code in the last 30 days and must have shipped a release in 2026. We re-run the numbers every month and log every change at the top of this page.