Agentes Regios

Agentes Regios is a real-time affective companion that turns raw EEG into a generative UI driven by an agent portfolio. A Master & Dynamic MW75 Neuro headset streams EEG at 4 Hz; our Python pipeline (neurable_connector + pidview) extracts an 8-dimensional affective signal — focus, stress, valence, arousal, joy, calm, excitement, neutral — as per-session z-scores. A contextual bandit (LinUCB) selects one of seven LangChain personas (breath_coach, body_scan, deep_focus, micro_break, reframe_cbt, env_tweak, social_nudge), and the chosen agent emits an A2UI action frame over WebSocket.

The interface goes far beyond chat. On macOS, a native Rust/Bevy/WGSL overlay (mood-orb) renders a generative blob avatar whose color, shape, distortion, speed and particles morph from the user’s mood vector in real time. When an agent fires, the renderer composes an A2UI compatible card — Card, Text, TextInput, Checklist, Timer, Button, ActionRow — directly over the blob. User interaction (clicks, submitted text, completed checklists, dismissed timers) flows back as feedback that closes the loop and trains the bandit, so the system learns which interventions actually help this user.

Stack: A2UI schema for the generative UI · LangChain for the agent portfolio · LinUCB contextual bandit · Anthropic Claude + Google Gemini + Ollama Cloud as interchangeable LLM backends · MW75 Neuro EEG as the affective input · Bevy + WGSL for the blob shader · bevy_egui for A2UI cards · WebSocket transport on 127.0.0.1:7777 · Rust + Python with uv.

We reused a NW 75 which we reimplemented for the project, but everything else is novel.