A provocative idea from software engineer Phillip Mortimer: as AI generates increasingly large amounts of software, code may become effectively “write-only.” The argument is that AI-generated implementations can become too dense or voluminous for humans to understand and review line by line. Instead of treating source code as the primary explanation of how a system behaves, developers may increasingly rely on tests, observability and other automated mechanisms to establish what the software is supposed to do.
Mortimer argues that this also makes code increasingly disposable. When an AI-generated implementation is difficult to debug or modify, it can be cheaper to preserve the tests that define its required behavior, discard the implementation and have an AI agent generate a replacement. In this model, the valuable artifact is not necessarily the code itself but the specification expressed through tests and system behavior.
The article also argues that human review becomes a bottleneck when AI can generate code at enormous volume. Mortimer recommends automated code-review agents running in CI, using separate instructions and context to challenge the generated code. He also proposes “de-slopify” agents that periodically inspect codebases for issues such as duplicated code, weak tests and unused feature flags. More ambitiously, coding agents could monitor observability systems and automatically create pull requests to address recurring production problems.
The deeper shift is that AI could separate software intent from implementation. Developers may spend less time manually expressing an idea in a particular programming language and more time defining what the system should accomplish, testing that behavior and improving the underlying product. Mortimer therefore sees creativity—not typing code—as the increasingly important engineering skill. But the approach also raises an important engineering question: if humans stop understanding much of the implementation, then tests, observability, architecture constraints and automated verification become critical safeguards against software that works today but becomes impossible to reason about tomorrow.
