Jin-Ho Baek

Native state management in agent runtimes solves problems that retrieval benchmarks cannot measure.

Majority voting among AI agents fails when they aren't truly independent.

Discovery is infrastructure, not afterthought, in always-on agent systems.

Classical assumptions break when agents produce distributions, not fixed outputs.

Agents fail from badly managed context, not small windows.

Fixed cloud-era scheduling breaks agent sessions, and runtime adaptation fixes it.