Why I built the HuggingFace for RL agents — and why RL needs one

If you've ever tried MineRL or OpenAI Five, you know the feeling.

The environment is fascinating. The problem is hard in all the right ways. And then you check the compute requirements — and close the tab.

RL has a compute problem. The most interesting environments are locked behind serious hardware. That means most people never get to play with the fun stuff. And it means that even if someone builds a great custom environment, there's no easy way for others to actually use it, train on it, or compete on it.

That's the gap I wanted to fix.

Introducing Agenlus — a browser-based RL training platform where you can train agents, share them via HuggingFace, and battle others on a global leaderboard. No install. No GPU bill. Just open your browser.

The goal isn't just accessibility. It's compounding knowledge — the same way HuggingFace made NLP and CV ecosystems compound on each other, Agenlus is built so RL agents and environments compound on each other.

I built this solo, and launched it this week. Would love feedback from the RL community on what environments you'd actually want to train on.

🔗