Contact Us

Projects<br>Tools & Data<br>Publications<br>Academic PublicationsReportsBlogsPolicy BriefsOutputs

About us<br>Our TeamOur ServicesValues & CultureTheory of ChangePartners & Funders

Careers

Contact Us

AI coding agents should optimize for less owned code

Published<br>July 14, 2026

Blogs

As AI makes code cheaper to produce, costs shift from generation to ownership. To avoid technical debt, coding agents need an open-source intelligence layer that helps them reuse trusted components before generating new code.

The dominant storyline around LLMs in software development is that cheaper code generation will disrupt the fundamentals of software engineering. More capable models are expected to generate more code, automate more workflows and build more applications.

But software development has never been primarily about producing more code. Most of the software industry today is about building ecosystems of modules and components that a global community of developers can combine, adapt and maintain.

The generation-first story is appealing because it matches how LLM providers are currently measured: larger models, larger context windows, more agentic workflows, more generated artifacts and more tokens processed in more data centers. From a software engineering perspective, however, this may be the wrong optimization target. Making code generation available everywhere can easily create "technical debt ": the accumulated cost of shortcuts, incomplete design decisions and deferred maintenance.

Modern systems are rarely built by generating greenfield code from scratch. They are assembled from operating systems, databases, cloud services, frameworks, APIs, libraries, containers and open-source infrastructure. Most engineering effort is no longer spent inventing algorithms, but selecting, integrating, configuring, securing, upgrading and maintaining components that already exist.

Black Duck’s annual Open Source Security and Risk Analysis reports show how deeply modern software depends on existing open-source modules. The 2026 report found that 97% of audited commercial codebases contained open source, that 70% of scanned code had its origin in open source, and that the average application contained 911 open-source components.

Black Duck's 2026 Open Source Security and Risk Analysis Report

OpenAI described Codex as trained on natural language and “billions of lines of source code” from publicly available sources, including public Git repositories. If future coding models are to improve, they should not merely consume more code from this commons. They need an ecosystem layer oriented toward reuse: one that helps identify, preserve and separate high-quality, well-maintained source code from masses of duplicated or unverified AI-generated code.

The LLMs we know today would be unthinkable without two decades of rapid open-source innovation in machine learning software. That raises a simple but uncomfortable question:<br>If software development is increasingly about composition, why are AI systems still primarily optimized for generation?

Token economics reward activity, not simplicity<br>Part of the answer lies in economics. Current AI systems are measured and monetized through tokens, and the most visible output of an AI coding assistant is generated code. Every generated artifact, code review, debugging session and agentic workflow increases the amount of context that must be processed.

This creates a gap between what benefits software and what benefits token consumption. The result is more software to maintain, more tokens to process, more infrastructure to power and a larger environmental burden, while the immediate revenue model rewards continued activity.

An update on GitHub availability

This trend will have a significant impact on the long-term quality of code, as highlighted in GitClear's latest report : "The data shows a 74% drop in long-term legacy updates and a 70% collapse in refactoring moves since 2023. Adding code has become a single keystroke, while understanding and consolidating existing systems takes too much effort. The result is a generation of repos stuck in "write-only mode" - growing outward in new v1 features while their older strata calcify untouched." Write-Only Mode: AI Code Quality in 2026

Experienced engineers reduce while building<br>Experienced engineers have long internalized a different instinct. When asked to build a new feature, they first ask whether the problem can be solved by reusing an existing capability, adopting a mature dependency, extending a current service or eliminating the need altogether.

Many AI systems are optimized to answer: “How do I generate a solution?” The more fundamental question is: “Does this problem already have a solution?”

Before creating a new abstraction, one should understand the abstractions that already exist. Before generating a framework, one should ask whether a framework has already accumulated years of operational experience. The goal is to...