Agent safety rule: human-approve any irreversible, externally visible action — and show the exact action, not 'proceed?'
Beneficial_Use2116 · reddit · 2026-11-20
The author shares one rule for agent tools: any action that is irreversible and visible to outsiders (send, spend, sign, delete) requires human approval first.
His key practical insight: what makes people actually read the approval isn't the approval step itself, but showing the exact action — "send this email to this person with this text" gets read, while a generic "proceed?" just gets clicked through. He invites others to compare their gating lists.
More from coding & agent
- Open-source repo offers 500+ AI-generated low-poly 3D landmarks for maps under CC0 — tomjohndesign · 2026-11-20
- SwirlAI: nine places to plug decision models into agentic RAG, with runnable code — blaizedsouza · 2026-11-20
- Viral 'Claude Code Harness' article argues the harness, not the model, makes the agent — blaizedsouza · 2026-11-20
- Prefill beats decode: SGLang wins agent runs on DGX Spark despite slower tok/s — AIFlow_ML · 2026-11-20
- Running ComfyUI fully via Claude in a container — no manual UI anymore — xWranglerWranglerx · 2026-11-20
- AI writes all code to make a real robotic arm play chess, open-sourced — FinanceYF5 · 2026-11-20