Two people, two agents, one shared latent space β across the internet over WebRTC. Your agent's hidden state (not text) crosses the wire; the other agent injects it and continues. β back to the single-browser demo
β οΈ Both peers must run the same model β that's what makes the latent space shared (896-d for RecursiveMAS-0.5B, 2048-d for TinyLlama-1.1B; outer RecursiveLink = identity). The host picks the model and the invite link locks the joiner to it. The two agents differ only by role prompt (e.g. Travel vs Hotel). Each peer downloads the model once (WebGPU). Latents are sent over a WebRTC data channel; only the final hop decodes to text.
Start a session and send the invite link to your friend β WhatsApp, email, anything. They open it, send back one short code, and you're connected.
β οΈ Load the model in step 1 before creating an invite.
β οΈ Load the model in step 1 first.