Mu uses Claude Opus 5.5 to build autonomous AI VTuber room

The implementation separates judgment from the body, using a JSON format to define room objects, their properties, and allowed actions. The character has 'eyes' that filter visible objects based on field of view and distance, with memory that blurs over time. Mu also added data-driven gimmicks, such as energy levels, object repair tasks, and event triggers like power outages. The character can interact with viewers and other AI visitors.

Mu noted significant latency, with one instruction taking 12–17 seconds and eye-only movement taking around 30 seconds. The character sometimes describes things it is not looking at or repeats greetings in multi-agent conversations. Mu proposed future improvements, including reducing LLM call frequency by using behavior trees for routine actions and adding object-to-object relationships to the room description.