Human-in-the-Loop

Definition. A required human review/approval step before an agent action takes effect. Serves two purposes: safety (a single mistimed touch can destroy relationship capital) and — less obviously — error interruption and data collection, since each human edit is a labeled correction the agent learns from. The checkpoint that breaks the compounding-error chain.

@misc{shalvi_gtm_human_in_the_loop,
  author = {Singh, Shalvi},
  title  = {Human-in-the-Loop — GTM World Model Glossary},
  year   = {2026},
  url     = {https://shalvisingh.com/gtm/glossary/human-in-the-loop}
}