TL;DR

I bring you:

• Sandboxing environment to train shell agents: repo
• A 14k shell task dataset for training on HF
• Synthetic data generation pipeline to generate customizable task datasets
• Batteries-included script to run RL with the above and your model of choice

Intro

Earlier this year I started messing more with “agents”, so much more that it ended up changing what I do at work (cool).

This Christmas I was thinking on revisiting a Rust port of croc I wrote some years ago to teach myself the language but ended up doing something different.

I had just been introduced to Cloudflare’s Durable Objects and thought of porting the port and check how hard it would be to build the same thing on the Cloudflare stack.

It proved to be incredibly educational and quite fun so I ended up using it in an introductory talk for the stack.