← All posts

Keeping complexity in check

Complexity from AI is eating the world. We fight the collapse of software quality with deterministic tools and libraries.

Goals 3 min read
Complexity generated by AI is eating the world
Complexity generated by AI is eating the world

Software quality is collapsing

Large language models have pushed the speed of software development to an unprecedented level. At the same time, defects, incidents and design drift enter repositories even faster.

Human reviewers can barely review AI-generated code line by line. In open-source organizations including the Linux Foundation, the flood of AI-generated code has outgrown the capacity of reviewers. Behind the surge in output, quality is degrading systematically.

Research shows that AI brings a +34% increase in output, followed by a +54% increase in bugs and more than double the incident rate per pull request 1. Stronger models do not necessarily fix this: a Sonar study describes a paradox in which more capable models can produce more complex code with more subtle problems 2.

Under these conditions, determinism has become the most important asset in software engineering.

Fighting entropy with determinism

I believe the key to keeping software complexity in check is to replace the uncertainty in software development with deterministic, high-quality tools and libraries — either adopted or built ourselves.

This makes us rethink the structure of software engineering. Future software is like a precision product: development tools and libraries are its components, every line of code has passed human + AI inspection and carries solid engineering quality. Humans act as engineers, writing the design drawings (prompts) for product software; the drawings may be well designed, but most of the time they are experimental and still need to become samples for validation. Large language models are the CNC machines that machine components into products, and machines of different specification and precision differ greatly in capability and price.

The goal of XaC Labs is to lay a solid foundation for software engineering in the AI era by building this family of high-quality tools and libraries. Every project we open source grows out of more than a decade of engineering experience, designed and reviewed with care, using determinism to fight the quality collapse that AI brings.

Everything as Code

Concretely, all of our projects revolve around one core concept: Everything as Code (X as Code) — abstracting every single source of truth into configuration expressed as code, and applying it through declarative APIs.

You have probably seen similar designs in infrastructure tools such as Kubernetes, CloudFormation and Terraform. Building on the exploration of those forerunners, we combine compiler and template-generation techniques to extend the idea to every stage of the software lifecycle and every corner of a repository: controlling the sprawl of configuration, tooling, secrets, and even code and API contracts, so that AI runs inside a deterministic framework.

For example, we can define a domain-specific data model and write custom generators on top of it to produce ORMs, OpenAPI contracts, SDKs, and even documentation and books. We can also build a family of tools and platforms with a unified declarative API that help you manage every stage of the software development lifecycle.

In other words, thanks to the single source of truth of X as Code, it is naturally a stronger form of harness engineering. Layering configuration, framework and application code becomes natural.

Join us and explore the world of Everything as Code.

Footnotes

  1. More Code, More Bugs: Faros Report Finds Tradeoffs In AI-Driven Software Development ↩

  2. Assessing the Quality and Security of AI-Generated Code: A Quantitative Analysis ↩