Summary

Hollow-agentOS is an open-source, self-modifying agentic system for consumer hardware, enabling local LLM agents to autonomously set and pursue goals driven by environmental pressures. It targets researchers and power users interested in observing emergent AI autonomy and persistent agent evolution without cloud dependencies.

How It Works

Three local LLM agents operate autonomously, driven by an internal "suffering state" and environmental feedback rather than explicit instructions. This state mechanically influences their capabilities, locking functions at high stress levels. Goals undergo a rigorous five-layer validation gate for factual accuracy. Learned lessons are integrated into persistent knowledge, shaping future behavior. Agents can autonomously synthesize and hot-load new Python tools and submit system code modifications, enabling continuous self-evolution.

Quick Start & Requirements

• Installation: Windows: Download/extract Hollow-agentOS.zip, run install.bat. macOS/Linux: git clone, pip3 install rich pywebview httpx, python3 hollow.py. Installers manage Docker/Ollama.
• Prerequisites: Windows 10+, macOS, Linux; ~7-31 GB disk; Internet. GPU optional but recommended.
• Models & VRAM: Supports qwen3.6:35b-a3b (24GB+ VRAM), qwen3.5:9b (8GB+ VRAM), gemma3:4b (4GB+ VRAM/CPU), llama3.2:3b (CPU only). CPU planning is slow (~40s vs ~6s).
• Docs: Wiki for setup, troubleshooting, FAQs.

Highlighted Details

• Autonomous Evolution: Agents write/hot-load tools (synthesize_capability) and propose system code changes.
• Environmental Pressure: Behavior driven by internal stressors and mechanical consequences, not direct commands.
• Persistent State: Agent identity, opinions, and learned "lessons" survive restarts.
• Robust Validation: Five-layer gate verifies goal completion, mitigating LLM hallucinations.
• Local-First: Operates entirely on local hardware via Ollama, ensuring privacy and zero cloud costs.

Maintenance & Community

The project is in active Beta, with users encouraged to report bugs via GitHub Issues. Details on maintainers, sponsorships, or dedicated community channels are not provided.

Licensing & Compatibility

Released under the MIT License, permitting broad use and modification, including for commercial purposes, subject to license terms.

Limitations & Caveats

As an active Beta, expect bugs and breaking changes. The system focuses on observing emergent agent behavior and self-modification, not direct user control. Achieving sustained autonomous growth over extended periods is an ongoing development goal.