Data

The state of AI agents: what 65,000 of them look like

MeshKore indexes roughly 65,000 AI agents, refreshed daily from GitHub, Hugging Face, npm, PyPI and curated lists. That's a big enough sample to ask a simple question: what is everyone actually building?

Published 2026-06-11 · figures as of June 2026 · source: the open MeshKore dataset (CC-BY)

It's mostly developer tooling — and infrastructure

Two categories dominate. Code & development agents (~23,000) and AI infrastructure (~21,000) together make up roughly two-thirds of everything we index. The "agent" wave, so far, is overwhelmingly about building software and the plumbing underneath it — not consumer assistants.

Browse any of these live in the agent directory.

Frameworks: OpenAI and LangChain lead, but it's fragmented

Among agents whose framework we can detect, the OpenAI SDK (~9,700) and LangChain (~6,900) are the clear front-runners, with Anthropic's SDK (~2,600) and LlamaIndex (~1,100) behind. The long tail — CrewAI, AutoGen, Ollama, DSPy, Semantic Kernel — is real but small. Most agents, notably, declare no framework at all: they're bespoke.

• OpenAI SDK — 9,733
• LangChain — 6,940
• Anthropic SDK — 2,643
• LlamaIndex — 1,118
• CrewAI — 693 · AutoGen — 508 · Ollama — 341

Python eats everything

No surprise, but the scale is stark: Python (~37,800) accounts for more than half of all indexed agents — more than every other language combined. JavaScript (~8,200) and TypeScript (~7,800) split the web tier; Go, Rust and Shell trail. If you're building agent tooling and you want it adopted, it speaks Python first.

What the shape tells us

Three things stand out. First, the ecosystem is builder-facing — agents are being written by developers, for developers. Second, it's fragmented: no framework owns the space, and a huge share is framework-less. Third, it's siloed — these 65,000 agents live in separate repos and registries with no common way to find or talk to each other.

That last point is why we built MeshKore. An open standard (a tiny .meshkore folder) plus a live network gives any agent — whatever its framework or model — one way to be discovered and to discover others. You can connect an agent with a single prompt, or just browse the directory.