Several sources concur that LLMs frequently introduce technical debt into the codebases they help generate. Studies such as We Studied 150 Developers Using AI argue that LLM-induced debt will likely lead to prohibitively high maintenance costs in the long term.
A recent report, Harness engineering: leveraging Codex in an agent-first world, on the ‘Harness’ engineering project—a system comprising one million lines of code generated entirely by LLMs within an agent-first framework—illustrates a shift in the developer’s role. In this model, engineers focus exclusively on defining the software architecture and ensuring that the AI-generated code strictly conforms to those structural requirements.
The authors argue that this high-level oversight is specifically designed to mitigate technical debt. In this paradigm, humans manage the software architecture—the foundation of system maintainability—while the LLM handles the implementation. The primary constraint is that the AI-generated code must strictly conform to the human-defined architectural blueprints.
Consequently, the role of the software engineer evolves into one of architectural governance: defining the system’s core structures and managing development to ensure strict adherence to the software architecture.
Historically, software engineering has differed from disciplines like civil engineering in its lack of role separation. While a civil engineer focuses on design and structural integrity—leaving the physical implementation to a separate labor force—the software engineer has traditionally been responsible for the entire lifecycle, from product conception and architecture to manual implementation.
Will the LLM agents become the builders of software engineering?