Why not regulate the harness (software) instead of controlling model inference?

TraditionalWait9150 · reddit · 2026-09-20

A Reddit user argues AI regulation should target the harness (tool-call software layer) rather than model inference. Since agentic AI is a loop of inference → tool calls executed by the harness → results fed back, the actual 'actions' happen at the harness level. By analogy to nmap or bash — where laws punish misuse rather than ban tools — they propose holding harness builders accountable: mandatory human approval per command and built-in hard rule engines (e.g., Drools) to block malicious commands. Benefits: safety even if alignment fails, humans stay as final arbiters, and clear accountability for AI actions.

Original post →

More from AGI Musings

AGI Musings channel →