Docker-based virtual machine system. Each AI agent runs in an isolated container (‘cartridge’). A host orchestrator manages container lifecycle, inter-agent communication via shared volumes and network bridges, and resource allocation across the swarm.
Language models loaded from HuggingFace or local weights. No training data — inference only.
Successfully orchestrates multiple LLM containers on a single machine with isolated environments and shared communication channels.
Performance bounded by host machine GPU/RAM. Container startup adds latency compared to native process spawning. No distributed multi-node support yet.
