Why code in 2026? The real job after AI

In plain terms. The question is no longer "will AI replace devs" but "what's the point of coding when AI generates 70% of the text". Doug Turnbull's answer: coding remains the only way to make an intention executable, testable, versionable. What changes is where we put the value.

Context

The post "Why Write Code in 2026" (softwaredoug.com, July 9) is circulating on HN because it hits outside the usual narrative. The author - a long-time ML engineer - does not defend the heroic coder or the future "no-code". He presents a sharper thesis: AI has made prototyping free; it's the deployment that remains expensive.

Analysis

What code historically provided, and that AI does not replace:

• an executable intention - the only artifact that the machine can really execute remains code, not a prompt;
• a test surface - a model that "produces roughly" does not survive the diff on 200 files;
• a versionable trace - what the AI "decided" on a Tuesday evening must be able to be audited a month later.

The value shift goes towards three areas: the boundaries (what your system is worth on the data side, API contracts, invariants), the context (structuring a repository so that an agent - human or AI - does not get lost) and the review (knowing how to say "this diff does not make sense" when it seems to).

What this changes on the craft side

Three practical consequences:

1. Tests become the main editorial artifact: they describe the intention for the machine and for the next AI.
2. Modularity gains even more value - not for reuse, but because a well-cut module fits in a context window.
3. Naming becomes central again: an agent hesitating between charge and charge_amount will make a mistake that a distracted human would not have made.

Scenarios

Base (60%): Teams that treat AI as a "fast intern, to be framed" gain 20-30% velocity without degrading debt. Optimistic (25%): The best ones push the automation of refactoring and recover 50% - but only on already clean codebases. Pessimistic (15%): Teams that let AI write without structure see debt explode in 12-18 months.

Risks

The real risk is not replacement, it's the erosion of discernment - the less you write, the less you read, the less you know when a diff smells bad.

So what

Coding in 2026 is no longer producing text: it's framing, testing, arbitrating. The question to ask any team: how much time do you spend understanding what the AI wrote yesterday? If the answer is "we don't do that anymore", you're selling your future for present velocity gains.