AI engineers today often discuss loop engineering, a concept popularized by Boris Cherny, the Anthropic engineer behind Claude Code, who remarked that he has largely moved beyond prompting; instead, he writes loops that prompt the model until a condition is met. For a coding agent with a test suite to validate against, this approach is effective. However, for an enterprise RAG pipeline, the loop itself requires deliberate engineering, as it must handle mundane failures: the parser flattens the one table containing the answer, retrieval returns the page adjacent to the correct one, the model outputs JSON that fails schema validation, or the API times out mid-batch. The four bricks handle the clean path; loop engineering governs every other path, enabling the system to recover rather than spin.
A one-shot pipeline commits to its first attempt: parse once, retrieve once, generate once, and return whatever emerges. When retrieval yields nothing or the answer is incomplete, there is no second chance. A loop provides that opportunity: detect the miss, adjust, and rerun the weak step before the user ever sees the result.
This article serves as a standalone companion to Enterprise Document Intelligence, a series that builds an enterprise RAG system from four bricks. It sits alongside Article 7bis (context engineering for single-document RAG) and addresses the layer just above it.
Article 13 (the composite pipeline) glued the bricks together and introduced the dispatcher. This article zooms into what the dispatcher does when a single call is insufficient. Volume 4 explores the agentic version of the same discipline, while this piece stays at single-document scope. Here is the map: what loop engineering is, the anatomy of one loop (trigger, termination, recovery), the two scales at which it operates, failure modes to guard against, and the outer loop that humans control.
π§ New to the series? Start with the map: Prompt, Context, Loop outlines the three engineering layers every RAG system relies onβthe prompt (the call itself), the context (what fills the model's window), and the loop (when the next call fires and when it stops)βand walks through the entire series through that lens, article by article. It is the fastest way to see what is covered and where this article fits.
π The runnable companion demonstrates the loops in action: you trigger llm_parse with a simulated timeout and watch the backoff schedule count down, then flip complete_answer_found to false on a listing answer and observe the dispatcher widen the retrieval scope and regenerate. On GitHub: doc-intel/notebooks-vol1.
1. What Loop Engineering Is
1.1 The Third Layer of the Agent Stack
Three disciplines stack atop an LLM call. Prompt engineering writes the call: system message, user message, schema. Context engineering chooses what enters and exits the model's context window between calls (see Article 7bis). Loop engineering decides when the next call happens, what triggers it, when the loop stops, and how the system recovers when something goes wrong.
A polished prompt with clean retrieval upstream can still fail; a loop adds resilience by allowing the system to retry, adjust, and improve. As of 2026, loop engineering has become a critical discipline in enterprise AI, with tools like LangChain and LlamaIndex offering built-in retry logic, but the underlying principles remain the same: treat the pipeline as a state machine, not a series of one-shot calls.
[The original content continues here; due to the truncation, the remaining sections (anatomy of a loop, small vs. big loops, failure modes, outer loop) would be included as per the source.]