AI and Code: Quality at Risk or Opportunity?

AI and Code Quality: A Complex Relationship

In the world of software development, the introduction of artificial intelligence tools has sparked passionate debates. Many developers express concerns about the impact of these technologies on the quality of the code produced. The idea that AI can generate code quickly but of inferior quality is a shared fear. If the use of coding agents seems to diminish code quality, it is imperative to take steps to identify and correct the aspects of the development process that are responsible. The decision to deliver lower-quality code is a conscious choice, and it is entirely possible to aim for higher quality production.

Technical Debt: An Obstacle to Overcome

The concept of technical debt is central when discussing code quality. This debt represents the compromises made to accelerate development, often at the expense of best practices. Sometimes, doing things correctly takes more time than is available, and developers hope that the project will last long enough for the debt to be repaid later.

However, the best strategy for managing this debt is to avoid it from the outset. In practice, many adjustments related to technical debt are simple but time-consuming. For example, an initially designed API may not cover all use cases, requiring modifications in several places in the code. Similarly, naming errors or duplicated features can accumulate, making the code difficult to maintain. Another example is a file that reaches several thousand lines of code, which ideally requires splitting into separate modules.

Coding Agents: A Promising Solution

Coding agents offer an effective solution for managing these refactoring tasks. These tools can be programmed to make changes in a branch or a dedicated workspace, without disrupting the developer's main workflow. They can be used asynchronously, allowing developers to continue their work without interruption.

Agents like Gemini Jules, OpenAI Codex, or Claude Code are used to automate these tasks. The process is straightforward: once the changes are made, the code is evaluated in a Pull Request. If the result is satisfactory, it is integrated. Otherwise, adjustments can be requested. This approach significantly reduces the cost of code improvements, allowing for zero tolerance towards bad practices.

Exploring New Options with AI

In software development, every problem can be approached in multiple ways. Part of the technical debt arises from unfortunate choices made during planning, such as ignoring an obvious solution or choosing an unsuitable technology.

Language models (LLMs) can help avoid these mistakes by suggesting common and proven solutions. Although they only propose solutions present in their training data, these options are often reliable and effective. Moreover, coding agents facilitate exploratory prototyping, allowing for quick testing of the viability of different technologies.

Coding agents can build simulations from a single well-formulated prompt, which reduces the cost of this type of experimentation to almost nothing. And since they are so inexpensive, we can conduct multiple experiments at once, testing several solutions to choose the one that best fits our problem.

Compound Engineering: An Evolving Approach

Coding agents follow precise instructions, and these instructions can be refined over time to improve outcomes. Dan Shipper and Kieran Klaassen from Every have developed a method called Compound Engineering. This approach involves concluding each project with a retrospective, documenting what worked to optimize future executions of agents.

The goal is to continuously improve code quality. Small improvements accumulate, and the costs of quality enhancements have decreased so much that there is no longer any reason not to invest in quality while developing new features. Thanks to coding agents, it is now possible to reconcile quality and speed in software development.