How AI actually changed day-to-day development in 2026 — and what didn’t budge

My mornings look different. My incidents do not. That is the honest split. In 2026 I start more days with a generated draft than with an empty buffer. I still end too many weeks in the same channel I ended them in 2019: a metric that moved the wrong way, a person who needs a sentence, a change I will not let a model merge because I will be the name on the page. The demos promised a new job. I got a new first hour and the same last hour.

I already argued why shipping still needs an engineer in the loop. This is the calendar version: what actually moved in the day, and what refused.

What changed: the blank page died

I used to spend the first forty minutes finding the files and writing the boring mapper. I now spend them checking whether the mapper the model wrote matches the sad path I care about. That is a real change. I ship more first drafts. I throw more first drafts away. The day has more material in it. It does not always have more finished work, because finished work still waits on review, on a flag, on someone in another timezone.

Search changed. I used to grep and hope. I now ask for the call sites and then I verify. When the index is good, I find the third file faster. When the repo is a mess, I get a confident tour of the wrong module. The day includes more “prove the tour.” That is new labor. It replaced “find the file.” I do not miss finding the file. I do not love proving the tour when I am tired.

Tests changed in a specific way. I get stubs. I still write or rewrite the one test that would have caught the last bug. If I skip that, I get a suite that loves the generated code and hates production. The day has more green. Green is not the change I needed. The change I needed is still the sad path.

What changed: I review more than I type

My output is more comments and more refusals. That was always part of senior work. It is now part of Tuesday for people who used to be measured on tickets closed. If your company still measures tickets closed, the day will feel like theft: you produced less “code” and more “no.” I have had to explain that no to a manager who still wanted a burn-down. The burn-down did not budge. The incident count did, when we refused the pretty diffs.

Pairing changed. I pair with a model and then I pair with a human for the part the model cannot see: the customer, the calendar, the thing we promised not to store. Human pairing got shorter and more about the radius. I like that when the other person is ready. I do not like it when pairing becomes “watch me accept.”

What did not budge: other people

Standups, design arguments, waiting on legal, waiting on a partner’s sandbox, waiting on a migrate window. AI did not move those. It made the wait more annoying because my buffer filled faster and then I sat on a finished draft for two days. The day now has a new mood: ready and blocked. I used to be blocked and not ready. I am not sure which I prefer. I know which one I blame the model for, wrongly.

Meetings did not get shorter because I can generate notes. Notes were never the length. The length was the decision. Models summarize. They do not decide. I still sit there. The calendar did not budge.

What did not budge: production

Production data is still weirder than staging. Restores still need a person. On-call still rings. The model will draft the query. It will not feel the lock. It will not remember the last time we ran this on a table that size. My incident days look like my 2022 incident days, plus a tab where I ask for a timeline I will edit. The tab is new. The sweat is not.

Deploys still need a window. Flags still need a name. Customers still need a sentence that is not a completion. I have used a model to draft the status page. I have still had to delete the sure paragraph. That edit is the job that did not budge.

What did not budge: knowing what we are building

Ambiguous tickets are still ambiguous. The model will pick an interpretation and build it. Faster wrong is still wrong. I spend the same time in product conversations I always did. Sometimes more, because a generated prototype makes a bad idea look done. Done-looking ideas are harder to kill. That is a new political problem. The need to kill them is old.

I thought AI would reduce the “what do you want” meetings. It increased the “no, not that” meetings. The day did not get quieter. It got more visual. Visual is not agreement.

How I structure a day now

Morning: generate and refuse while I am sharp. Midday: humans, because they are awake. Afternoon: the sad-path test and the radius note. Evening: I try not to merge anything the model loved if I am tired. Tired-me accepts. That did not budge either. The model made tired-me more dangerous because the PR looks finished.

I block time for review the way I used to block time for writing. If I do not, the drafts pile and the loop clogs. The company sees “AI productivity” and wonders why the board is full of open PRs. The board is full because generation is not the constraint. Review and windows are. Those did not budge.

What changed for juniors, and what did not

Juniors get a draft now. They used to get a stuck morning. That is kinder. They also get a finished-looking PR they cannot defend. That is less kind. The day-to-day change for them is a faster way to look done. The part that did not budge is the need to explain the radius. I spend more of my day on that explanation than I did when their stuck morning was visible. Invisible stuck is a generated file. I would rather see the stuck. I now ask them to narrate what they refused. If they refused nothing, we do the ticket again. That ritual is new. The need for a human who can refuse is not.

Mentoring hours did not shrink. They moved. I used to sit on syntax. I now sit on “why this hop exists.” Syntax budged. Systems did not. Anyone who says juniors can skip the loop because the model writes is describing a day I will not staff. I will staff a day where the junior owns a small noun and uses the model as a typist. The noun did not budge as the unit of the job. The typing did.

What I tell people who ask if the job is more fun

The first hour is more fun. The last hour is the same. Fun is not the metric I use for whether we staff fewer humans. The metric is whether the parts that did not budge still have names. If they do not, we have a faster typewriter and the same outage. I like the typewriter. I will not confuse it with a smaller incident. The incident did not read the launch blog. It read the metric. The metric did not budge until a person did.

Day-to-day development in 2026 is a faster start and an unchanged finish. I plan the day around that. I do not plan it around a demo where the finish was a screenshot. Screenshots budged. Production did not. I still work here, in the part that did not budge. That is the part I want paid for. The rest I will let the model have, until it starts looking like a merge. Then the day becomes 2019 again, on purpose. If a tool vendor asks what feature would change the finish, I tell them: none that stay in the editor. The finish lives in flags, windows, and people. Build those if you want the day to move. I will keep using the editor features. I will not pretend they moved the part of the day that pays for the rest of the company.