---
title: Loop Engineering Is Not a New Idea
date: 2026-07-01
summary: A short note on agent iteration, evidence, and human review.
source: https://x.com/WuDavidxx/status/2071488369373434221
---

When I first heard the phrase "Loop Engineering", it sounded like a new name for old methods. Automated research, automated tests, and review loops all do the same thing: look at the result, then decide the next round.

For agents the loop is short. A person sets the goal and the limits. The agent runs and keeps evidence. If it fails, it repairs or rolls back. When a stop condition is met, it hands a report back to the person.

> The point is not to run more rounds. It is to leave a basis for judgment after every round.

## The problem with one-shot answers

A polished first output is enough for a demo. In real work, research misses sources, code misses tests, and pages break on small screens. Without an inspection standard, an agent is only repeating generation in a different form.

So I write down what counts as done, which evidence must be kept, how many rounds are allowed, and when a person has to step in.

## How I use it now

[Pi Company](/works/pi-company/) manages several agents with tasks and review gates. [Pi Fusion](/works/pi-fusion/) keeps the participant models, the judge, and the evidence. [Vizual](/works/vizual/) puts the result into an interface you can keep working with. The three projects work differently, but each needs a chain of work you can inspect.