Gergely Orosz says cloud-hosted AI agents need deterministic controls for sensitive data

Software writer Gergely Orosz says cloud deployment will increasingly host AI agents but does not solve their risky handling of sensitive data.

He argues that because agents are nondeterministic, guaranteeing a harmful action never happens requires a deterministic system around them. His post does not describe which safeguards he has in mind.

A separate SiliconANGLE article says nondeterministic LLM-based agents can give different outputs for the same input, so they suit tasks with changing inputs rather than single correct answers. It says agents are a poor fit where deterministic software suffices, where every action must be exactly correct, where misbehavior could cause serious harm, or where cost exceeds expected benefit, and it recommends an error budget to decide whether to deploy one.