Research on how agents talk to each other below the level of text.
A null modem cable linked two PCs directly, with no network in between. This repo asks what agents can exchange when the harness owns the model, the tokenizer and the machine, and doesn't have to pass everything through text.
Agents today talk in tokens because hosted models only accept tokens. Exo running on exo-os changes that: the harness and a local model work together at the root, down to memory, caches and the tokenizer. Agents running that way across several devices could share far more than messages.
- Research notes and design write-ups.
- Small experiments on local open-weight models, such as sharing a KV cache between agents, passing embeddings or other latent messages, sharing a tokenizer, and keeping memory in sync across devices.
- Measurements against a text baseline (plain English and ABIOS), with raw logs and rerun commands.
Ideas that prove out here move into exo-os and exyber, one profile first and then outward after verification.
See CHARTER.md for scope, questions and success criteria.
MIT