The core idea is to map the agent's internal state to visual cues. I've implemented "Aura" (a white light that flashes to indicate what the LLM is focusing on before it replies) and environment triggers. For example, if a git push
succeeds, the character shows a happy expression and fireworks trigger; if the agent hits a crash or an API error, the character looks sad.
Technically, this works via a self-referencing MCP (Model Context Protocol). The AI agents actually control the lighting, the character's body, the scene, and the UI effects in real-time. I've also added extension support with hot re, making it a self-modifiable terminal similar to how Emacs operates.
If you want to test this AI workflow on macOS:
brew install --cask sktkkoo/yorishiro/yorishiro
It turns the abstract process of prompt engineering and agent execution into something tangible. Instead of staring at a blinking cursor while waiting for a complex task to finish, you get environmental notifications when the agent is waiting for your approval.
Next AGI and the End of Intelligence Scarcity →