The Claude Code team defines four categories of agentic loops distinguished by their trigger and stop conditions: turn-based loops with manual verification, goal-based loops (/goal) that iterate until success criteria are met, time-based loops (/loop and /schedule) for recurring or event-triggered tasks, and proactive routines combining these primitives with auto mode for unattended work like bug triage. The piece stresses managing token usage through clear scope boundaries, deliberate model selection, and deterministic verification steps encoded as reusable skills.
