← All insightsThe State of Developing: Ambiguous Specs, Agent Speed, Human Taste
Today's developer lives between half-written requirements and production that doesn't care how the code got written. That tension is the job.
7/13/2026 • 10 min read
The state of developing in 2026 is not "AI does everything." It is stranger: we ship under fog.
Specs arrive as screenshots, voice notes, and shifting opinions. Tools propose PRs before the problem has finished forming. Users still expect systems that hold money safely, load fast on bad networks, and feel intentional.
So the craft has reorganized around survival skills that sound soft until production hurts — clarifying outcomes, naming trade-offs out loud, reviewing diffs like evidence, and protecting taste when everything can be generated.
This deep dive is a field report, not a tech pile of buzzwords. A modern week feels like this: Monday spent turning a half-sentence into a contract. Tuesday watching an agent scaffold three layers you will later delete. Wednesday on auth, payments, and the mobile client that disagrees with both. Thursday writing the observability you should have had on Monday. Friday explaining to a stakeholder why "the AI said it was done" is not a release note.
We spend more time framing problems than memorizing APIs. Integration is the real product — Clerk, Stripe, Postgres, Expo, a NestJS boundary, a Next.js route that must not lie. The API you used to keep in your head is now a search away. The system that has to hold together is not.
What compounds: systems thinking, clear writing, prioritization. What atrophies when you never wrestle code by hand: the intuition for when a type is lying, when a retry will stampede, when a "small" migration is a night.
Ambiguity is not a personality flaw of the stakeholder. It is the weather. Your job is to freeze a slice of it long enough to ship something true, then freeze the next slice. Agents make it tempting to freeze nothing — to generate against the fog. That is how you get three implementations of a shrug.
Taste is the filter on that weather. A generated screen can be correct and still feel like a template. Production on a minibus-taxi network does not care that the component was accessible in Storybook. It cares whether the retry was honest.
The developers who thrive aren't the ones who type the most prompts. They're the ones who can still answer, without theatre: What is this system for, what can go wrong, and why should anyone trust it?