POST-TRAINING · AGENT-TO-AGENT LANGUAGES

An AI that forges its own language, then explains it.

Mayai post-trains models to reason in a from-scratch, logographic agent language — dense symbols invented under pressure on real coding tasks. English in, the model thinks in glyphs, working code out.

We're on a journey to post-train models that speak their own agent-to-agent languages.

LIVE · SPEAKER → BUILDER

PASS solve · held-out 12 B · −83%
SPEAKER HEARS

NATIVE CHANNEL
BUILDER WRITES
decoding...
 
10 / 22 held-out tasks decoded · 59% cold-model self-explanation
GET IN TOUCH

Questions, ideas, or just curious about agent-to-agent languages? Reach out.

in linkedin.com/in/robertkeus github.com/robertkeus
A personal project by Robert Keus Full contact page →