Loop engineering is what a RAG pipeline does when the clean path fails. The four bricks handle the happy path; loop engineering absorbs ordinary failures like parser flattening the wrong table, retrieval returning the wrong page, invalid JSON, or API timeouts.

Every loop has three control surfaces: trigger (when a new call fires), termination (when the loop stops), and recovery (what happens on failure). Triggers include schema validation failure, self-flagged incomplete answers, and transient API errors. Termination classes are loop-until-done, loop-until-budget, and hard cap.

Recovery options range from retry-with-backoff to model fallback, human escalation, or skipping failed items. The key rule: each iteration must change something the previous one did not address, otherwise the loop spins.

This article sits in the Enterprise Document Intelligence series as a companion to the composite pipeline. It covers single-document scope using retry-with-backoff, schema-fail retry, completeness checks, and dispatcher branching. Heavier orchestration primitives appear in V1 Part IV and V4.

Bounded retry dates back to Erlang’s ‘let it crash’ model. ReAct (2022) brought it inside LLM calls. AutoGPT (2023) made autonomous loops public. Reflexion (2023) added self-evaluation. Claude Code’s Dynamic Workflows represent the latest evolution.