Hollow AgentOS adds persistent suffering states

The interesting part here is not “sentience” theater, it’s that a feedback loop built around observable action changes agent behavior in ways a normal prompt never would.

• –The system treats state as something earned through actions, which is closer to control theory than chat prompting
• –A 9B model still hits obvious reliability limits: the post says roughly half the synthesized tools crash, so execution quality remains the bottleneck
• –The overnight run shows emergent coordination patterns, including similar stressor naming and repeated focus on the same files without direct prompting
• –The local/Ollama/MCP setup makes this a practical agent-runtime experiment, not just a speculative demo
• –The biggest takeaway is that persistent memory plus stateful incentives can meaningfully steer small models, even if the outputs are messy