Xenolect

README excerpt

# An RL enviorment to train the ability to design communication protocols for terse agent to agent communication

Agent-to-agent communication will surpass human-to-human communication in the future. But right now, agents aren't great at using English — and we think we can do a lot better.

---

[![Watch the overview](https://img.youtube.com/vi/D6Q5o14w3ks/0.jpg)](https://youtu.be/D6Q5o14w3ks)

---

## The Problem

Here's a meeting scheduling task in plain English:

> *"I'm available Monday morning in San Francisco, from 9 to noon. Are you free then?"*
> *"Unfortunately I'm in New York on Monday. How about Tuesday?"*

They waste time on pleasantries. "Unfortunately." Filler. Human language carries centuries of social overhead that agents don't need.

## The Idea

We believe there will be a different compact language for every different task. Rather than forcing agents to communicate in English, give them the ability to come up with an **alien language** — one invented for the task at hand.

Crucially, this isn't just a prompt engineering trick. We built an **environment specifically to train this capability into models** — so that inventing efficient communication protocols becomes a learned skill, not a one-off hack.

The agent generates a language specification. The environment evaluates it. Two agents then communicate using that spec to complete the task, and the reward is simply: how many tokens did it take?

Train on that signal, and the language gets better.

## What Ha