The result is a model-specific REPL "genesis snippet." It's a short bash script that prompts for your API key, asks the model to generate a REPL with an agentic tool-use loop, and then drops you straight into that running REPL.
The self-imposed rules:
1. No copy-pasting code after the initial snippet.
2. No manual editing—the agent has to write and fix everything for you.
3. Zero agent frameworks or libraries.
From there, the goal is to ask the agent to rewrite its own REPL, add quality-of-life improvements, and eventually figure out how to connect itself to Telegram.I put the snippets up on a site along with some open challenges like code golf and speed runs:
To try it, you just need your own API key and a Linux VM or Docker container. If you want to see how it actually behaves before running it, here is the raw terminal output of my first successful run: