AI Agents Are Writing More Code Than Humans — But Is Anyone Asking What Code Actually Is?

Breaking: The Nature of Code Under Scrutiny as AI Takes Over Development

In an urgent shift that is reshaping the software industry, artificial intelligence agents now write a majority of new code in many projects, raising a fundamental question: Will traditional source code even exist in the future? Experts warn that without a clear understanding of what code truly is, the entire profession of software engineering risks losing its intellectual foundation.

"We are delegating the act of writing code to LLM-based agents at an increasing pace, but we rarely pause to ask what we are giving away," says Unmesh Joshi, a leading software thinker and author of the influential essay on code's dual nature. "If we forget that code is not just a set of instructions but also a conceptual model of a problem domain, we may end up with systems no one can reason about."

The Two Purposes of Code: Instructions and Model

Joshi argues that code serves two intertwined roles. First, it tells a machine exactly what to do, step by step. This is the mechanical, executable layer. Second, it encodes a conceptual model of the real-world problem the software is meant to solve — the vocabulary of business rules, user behaviors, and domain logic.

"Most developers focus on the machine-instruction part, but the true value of code lies in the mental map it provides," Joshi explains. "When an AI generates code without a deliberate effort to preserve that map, the human team loses its ability to understand why decisions were made — a recipe for catastrophic failures in production."

Background: The Rise of AI Coders

Over the past three years, tools like GitHub Copilot, Codex, and Claude Code have moved from novelty to necessity. A 2024 industry survey found that 64% of developers rely on AI to generate at least half of their code. Startups now operate with tiny teams, with agents handling everything from boilerplate to complex logic.

Yet this trend has outpaced understanding. The original concept of "source code" as a human-readable blueprint is being blurred. If AI writes code that humans barely inspect, does that code still serve as a communication tool between developers? Or does it become an opaque binary? Joshi warns that without a shared language between human and AI, software projects will turn into unmaintainable black boxes.

What This Means: The Future of Programming Depends on Intent

The implications are stark. If code loses its role as a conceptual model, developers will become helpless troubleshooters, unable to modify or extend AI-generated logic. "We will end up with software that works until it doesn't, and then no one can fix it," Joshi says. "That is not a sustainable future."

To counter this, Joshi advocates for a new discipline: intentional programming. Teams must explicitly document the domain model and ensure that AI agents are constrained to preserve it. This means writing prompts that capture the "why" behind the code, not just the "what." It also means investing in code readability and architecture review even when machines do the typing.

• Key takeaway 1: Code is both a machine instruction set and a human knowledge artifact. AI erodes the latter if left unchecked.
• Key takeaway 2: Future developers will need new skills: prompt engineering for domain models, and the ability to read AI-generated code critically.
• Key takeaway 3: Companies should adopt guidelines that require conceptual transparency in AI-assisted workflows, with regular human reviews of the model layer.

A Urgent Call for the Industry

Joshi's analysis is a rallying cry. As AI agents accelerate, the industry must deliberately preserve the conceptual clarity that code provides. "We need to teach LLMs to think in terms of the problem domain, not just syntax," he says. "Otherwise, we will drown in code we no longer understand."

The question of what code is — once purely academic — has become an existential challenge for every engineer, manager, and product owner. The answer may determine whether the software of the future is robust or fragile.

This article is based on the original work by Unmesh Joshi, repurposed as breaking news. For a deeper dive, read the full essay: "What is Code" (link).