The Hollow AgentOS project provides a local, autonomous multi-agent framework using Qwen 3.5:9B models that manage their own goals, tool synthesis, and psychological states without cloud dependencies.

Key Points

• The system runs locally via Docker and Ollama, featuring 91 available tools and 128+ natural language-installable capabilities.
• Agents experience "suffering" states based on six stressor types, requiring measurable behavioral changes to resolve rather than simple text-based responses.
• Core infrastructure includes an audit kernel, VRAM-aware scheduling, atomic multi-agent transactions, and semantic memory using vector embeddings.
• Agents can synthesize new Python tools at runtime and request core system modifications through a formal specification queue for human approval.
• The framework supports persistent agent identity, checkpointing, and replayability to ensure state survival across system restarts.

Why it Matters

This project shifts the paradigm of AI development from building static applications to observing autonomous, self-governing agent ecosystems. By integrating psychological stressors and formal audit kernels, it provides a unique environment for studying long-term agent behavior and complex, multi-agent coordination in a controlled, local setting.