The Maintenance Cliff: LLMs and the Learning Loop

The Assembly Line Fallacy

In our industry, there is a persistent temptation to treat software engineering like an assembly line. We constantly aim to abstract work into higher-level languages, reusable components, and “black box” services to increase velocity.

In a recent article on MartinFowler.com, Unmesh Joshi critiques this tendency. He argues that this efficiency-obsessed approach clashes with the most fundamental property of our work: learning.

When we rely entirely on readymade solutions without understanding the “why” and “how” underneath, we treat code as a static asset rather than a dynamic expression of knowledge.

The Three-Phase Learning Cycle

Joshi outlines the process of acquiring deep technical expertise as a three-phase cycle:

1. Ingest: Absorbing information from various mediums (docs, books, tutorials) into the mind.
2. Practice: Applying and experimenting with that information in real contexts to see where it breaks.
3. Iterate: Repeating the practice phase across different scopes to deepen understanding and intuition.

The danger lies in skipping steps. Relying on reusable solutions without deeply engaging in this learning loop creates an illusion of competence. It provides an initial boost in velocity, but ultimately leads to high costs when changes are needed.

The code becomes a black box—difficult or impossible to modify because the maintainer lacks the context required to change it safely.

The LLM Accelerator (and Trap)

This is where Large Language Models (LLMs) complicate the equation.

LLMs are the ultimate “readymade solution” generators. They amplify the risk of bypassing the learning loop because they allow us to generate working solutions without ever grappling with the underlying concepts or language specifics.

While LLMs are incredible tools for brainstorming and bridging the gap between an abstract idea and a concrete implementation (like translating API specs into code), they cannot replace the essential struggle of learning.

The Verdict: There are no shortcuts around the need to practice and internalize knowledge.

The Maintenance Cliff

True expertise is built by learning and applying knowledge to build deep context. Any tool that offers a solution without this journey presents a hidden danger.

By offering seemingly perfect code at lightning speed, LLMs represent the ultimate version of the Maintenance Cliff: a tempting shortcut that bypasses the essential learning required to build robust, maintainable systems for the long term.

If you want to understand the real challenges of engineering growth in the AI era, Joshi’s article is a must-read.¹