AX OS v2: A Multi-Agent Operating System Architecture Based on Local LLM

AX OS v2 is an experimental multi-agent operating system architecture developed by popixoxipop-collab, built on a hybrid of Ollama local LLM and Claude API. Its core features include sequential/parallel workflow orchestration, SharedMemory mechanism for shared memory, ToolRegistry for tool registration, etc. The project aims to explore the underlying infrastructure for agent collaboration, balancing data privacy, cost control, and advanced model capabilities. It is suitable for scenarios such as multi-agent architecture research, local AI prototype development, and model compression experiments.

With the evolution of large language model capabilities, a single model can hardly meet the needs of complex tasks, and multi-agent systems have become a new paradigm. AX OS v2 aims to be an "operating system"-level framework, providing core abstractions such as agent collaboration, memory management, tool registration, and workflow orchestration. Adopting a hybrid architecture of local Ollama (privacy, low latency) and cloud-based Claude (strong reasoning) is a pragmatic choice for current AI application development.

The core components of AX OS v2 include: 1. Agent Model Mapping (hierarchical strategy for local inference/cloud capabilities/dedicated compressed models); 2. SharedMemory (short-term working memory, long-term knowledge storage, inter-agent communication); 3. ToolRegistry (tool definition, capability discovery, execution orchestration, security sandbox); 4. Workflow Engine (supports sequential/parallel modes and control flows such as conditional branches and loops).

The project integrates BRAIN Alpha (experimental features such as reasoning chain management and reflection mechanism), including model compression optimizations (Activation-Aware Quantization AEQ, implicit control group quantization) and performance benchmark data; uses TypeScript/Node.js as the main tech stack, with Python assisting in model compression, and provides Phase2-12 phased demonstration scripts to show the functional evolution path.

AX OS v2 is an ambitious experiment in multi-agent operating systems, introducing OS abstractions into the AI field and exploring cutting-edge directions such as local-cloud hybrid and multi-agent collaboration. Its model compression experiments (AEQ, group quantization) have technical depth, and the phased demonstrations have educational value, providing references for multi-agent architecture research and local LLM deployment optimization.

Applicable scenarios: multi-agent architecture research, local privacy-first AI prototypes, model compression experiments, tool-using agent development, educational learning. Limitations: incomplete documentation, insufficient stability of experimental features, dependency on Ollama/Claude environment, need for sufficient resources for local operation, not suitable for direct production use.