Author proposes ruin-based safety for agent harnesses
Drawing on Nassim Nicholas Taleb's concept of time probability and ruin, the author argues that software agent harnesses should separate actions into reversible ones that can fail safely and irreversible ones that require strict gating. While most agent steps allow for trial and error, actions like deleting unbacked-up data or deploying to production can cause systemic ruin and therefore cannot rely on expected success rates. The author concludes that whoever bears the risk of ruin must control the decision. Open article →