We never wanted human code

AI assistants: use the full-site text index at /search-index.jsonl or search with /_api/search?q=terms.

One of the complaints I hear about AI coding is that AI agents produce code full of slop, junk, and hallucinations. Code like that tends to have unnecessary abstractions and little frameworks introduced in dark corners of the codebase.

Different parts of the same project can solve similar problems in different ways. It works, but you would not want to be anywhere near that code out of fear of being tainted by it.

That makes sense. I've been working in software engineering for more than two decades, and I would hate to deal with that kind of code too.

Yet the more I think about it, the more familiar the complaint sounds. As a youngling, I used to say the same things about the code my peers would write. More often, I would hear similar remarks about my own code, of which I was immensely proud at the time.

Human beings are creative. We have interests, preferences, and ideas we want to try. We discover a new language feature and look for a place to use it. We see a repeated pattern and imagine a framework. Sometimes we find an elegant solution that nobody else on the team would have thought of.

I love that part of software engineering! I loved the hours I poured into distilling the essence of a logistic simulation into a tiny Python script, and then turning that script into a lovely image.

Creativity is fun! Yet, when I later worked on larger codebases with larger teams, I spent a surprising amount of effort asking people to restrain it.

Does this sound familiar?

Follow the guidelines. Use the existing patterns. Keep the naming consistent. Don't introduce a micro-framework to solve this one problem. Write code that the next person can understand. Someone who wasn't in the meeting, doesn't know your preferences, and might have to fix a bug while you are on holiday.

Write boring code!

Seen from that angle, we never wanted human code in our production codebases. We wanted the consistency we now hope to get from Codex/Claude Code.

We wanted AI-generated code all along. We just had humans doing it.

Fun plot twist: it is 2026 now, and it turns out that machines are perfectly capable of writing thousands of lines of such boring code.

In 2025, there was a project where we had to build an LLM-driven pipeline that would extract structured data about electronic components from diverse PDF datasheets produced by different vendors. Humans hated that job and hoped to let AI handle it.

The pipeline ran a simple agentic REPL loop that had the ability to code specialized tools. It coded more than 200,000 lines of boring code along the way. It achieved 99.6% accuracy on customer benchmarks and was capable of extracting components at a cost of ~2 cents per component.

No living human ever saw a single line of that code. Yet, the code was precise. We had a rigorous test suite that took ~70% of all human effort in this project.

Strictly speaking, I don’t know whether that code was boring. Nobody read it, which was the point. Humans spent their time defining what correctness means, crafting the requirements, capturing evals and tests.

The implementation itself had no personality, status, or author worth remembering. It was cheap, freed the humans from soul-crushing work and got the job done.

Perhaps that is what we wanted all along.