Software Engineers Are Writing Code They'll Never See

There's a peculiar dynamic emerging in software development that hasn't gotten nearly enough attention: engineers are increasingly shipping code they didn't write and may not fully comprehend.

Consider the recent announcements from Nextdoor and Endava. Nextdoor engineers are using Codex with GPT-5.5 to "investigate hard-to-reproduce issues" and "build across platforms." Endava is "redesigning software delivery around AI agents," implementing ChatGPT Enterprise and Codex to "automate development workflows." These aren't fringe experiments—they're production systems at companies serving millions of users.

The language both companies use is revealing. They talk about "building without limits" and "streamlining delivery processes." What they don't talk about is what happens when the engineer who shipped a feature generated by an AI agent leaves the company, and nobody on the team actually understands how that code works.

This isn't theoretical. We're already seeing the downstream effects in unexpected places. Federal courts report that AI-generated legal filings jumped from 1% in 2023 to 18% in 2026—and while judges note these documents are often "clearer and easier to understand," that's because AI tends to produce generic, template-driven content. The same dynamic applies to code: it may compile cleanly and pass tests, but lack the contextual understanding that comes from human design decisions.

The robotics world offers a useful parallel. When Amazon announced that its Proteus robot can now understand natural-language commands, the stated benefit was that "warehouse workers can issue commands without programming experience." But there's an implicit trade-off: those workers also won't develop programming intuition. They become operators of systems they can't modify or debug.

The pattern is consistent across domains. Generalist AI raised $400 million to build foundation models that achieve "99% success rates on tasks where previous models failed." That's impressive—until you consider what happens during the 1% failure case when no human on the team understands the underlying model architecture.

None of this means AI coding assistants are bad. They're extraordinarily useful, and the productivity gains are real. But we're building a technical debt of a different kind: comprehension debt. Every time an engineer uses AI to build across platforms they don't deeply understand, or investigate bugs in code they didn't write, we're creating systems that may work perfectly—until they don't.

The robotics and AI community has spent considerable energy on questions of safety, alignment, and verification. We scrutinize whether surgical robots can safely operate autonomously, whether counter-drone systems make reliable targeting decisions, whether foundation models can ground themselves in physical reality.

Perhaps it's time to apply the same rigor to the code generating those very systems. When the engineers building the AI are themselves relying on AI to write their code, we've created a recursive dependency that deserves more than a productivity celebration.

The next generation of developers may be extraordinarily efficient. But efficiency and understanding aren't the same thing. And in a world where AI agents are writing the code that controls robots, manages critical infrastructure, and makes autonomous decisions, understanding might matter more than we think.