Writing code stopped being the job

In 2019 I had Stack Overflow open in three tabs by default. One for the question I was avoiding, one for a half-remembered thread I was pretty sure had the answer, and one as a backup in case the first two led nowhere. The site was the developer’s external memory. Half my early projects were built on copy-pasted snippets that had been peer-reviewed by fourteen thousand strangers and a vote system I trusted more than my own instinct.

That site is not the site I remember. Stack Overflow has lost roughly 76% of its monthly question volume since ChatGPT shipped in November 2022. The community-vote, copy-paste, “marked as duplicate” loop that taught a generation of developers how to write code has been quietly replaced by a chat window. I cannot remember the last time I opened a Stack Overflow tab on purpose. When I land there now it is by accident, through a Google result that surfaces a 2017 answer that no longer compiles.

I started writing code professionally at 19. Most of it shipped to clients who paid by milestone, which is the kind of pressure that teaches you very fast which questions are worth googling and which are worth typing. I have spent the last six years inside that loop. Watching it dissolve in real time is one of the stranger things I have lived through.

Stack Overflow did not lose. It became redundant.

The decline is not a brand problem. It is a workflow problem. Asking a question on Stack Overflow used to mean writing a minimal reproducible example, formatting your code, picking the right tags, then waiting an unknown number of hours for an answer that might be helpful or might be a passive-aggressive duplicate flag. The unit cost was twenty minutes minimum. Asking a question of Claude or Gemini is one sentence and a paste. The unit cost is six seconds.

Once the cost of an answer drops by three orders of magnitude, the equilibrium changes. Nothing else needed to happen. The site is not being killed by a competitor. It is being killed by the fact that nobody needs it for the original purpose, and a Q&A archive does not survive long without people actively asking questions.

Google quietly redefined what engineering is

In April 2026, Sundar Pichai said in a Cloud Next blog post that 75% of all new code at Google is now AI-generated and approved by engineers, up from 50% in late 2025. He framed it not as a productivity stat but as a role change. “We’re now shifting to truly agentic workflows,” he wrote. “Our engineers are orchestrating fully autonomous digital task forces, firing off agents, and accomplishing incredible things.” A code migration that used to take a team months was completed six times faster by agents working alongside their human reviewers.

The interesting word in that sentence is orchestrating. The job description used to be “writes code.” The job description at Google is now closer to “directs the things that write code.” The Software Engineer title is becoming, in everything but name, a Product Architect title. You design the shape of the work and you decide what ships. The agents do the typing.

This is not a forecast. This is what is shipping in the largest engineering organisation on the planet, right now, in production.

Anthropic published their own numbers

Anthropic ran the experiment on themselves. In late 2025 they surveyed 132 engineers and analysed roughly 200,000 Claude Code sessions inside their own building. Employees self-report using Claude in 60% of their work and report a 50% productivity boost, a two-to-three-times jump from the previous year. Code design and architectural planning grew from 1% to 10% of their daily work. Implementing new features rose from 14% to 37%.

The number that actually matters is buried further down. 27% of the Claude-assisted work would not have been done at all without the model. Tasks that were too complex or too time-consuming to justify the manual effort. A nice-to-have internal dashboard. A migration that would have been deferred indefinitely. An exploratory build that would have failed the cost-benefit check the day it was proposed. The bar for “worth doing” has been quietly relocated.

Claude Code itself has gone full-loop. It reads the codebase, plans a sequence of actions, runs bash and git and the test suite, and if the tests fail it adjusts its plan and tries again. Anthropic has built a sandboxed Auto Mode where Claude works without asking permission for routine actions, with an ML classifier sitting on top that only stops the agent when it spots something high-risk: mass file deletion, data exfiltration, malicious code execution, an action that would irreversibly change system state. The classifier has a 0.4% false positive rate and a 5.7% false negative rate on synthetic exfiltration attempts. Engineers are trained to sit at the top of what one internal write-up calls a principal hierarchy. The humans are reviewers-in-chief. The agents work freely below. The human only steps in when the classifier flags something.

{
  "permissions": {
    "allow": ["Read", "Edit", "Grep", "Bash(npm run test:*)"],
    "deny": ["Bash(rm -rf *)", "Bash(git push --force *)"],
    "ask": ["Bash(*)", "WebFetch"]
  }
}

That is what writing code looks like now. You are not typing functions. You are defining a permission boundary and reviewing what comes out of the other side. The job is closer to running a small intelligence service than to authoring a textbook.

The engineer of 2026 is not the person who can write a binary search from memory. It is the person who knows when to merge the pull request and when to send it back.

The new half-life

There used to be a comfortable rhythm to staying current. You could coast on the same stack for a decade and remain employable. Java in 2008 was recognisably Java in 2018. The half-life of a useful skill was measured in years, sometimes whole careers.

That window has closed. The half-life now is three to four months. A model lands, a CLI lands, an agentic workflow lands, and the people who picked it up in week one are already two cycles ahead of the people who waited for the dust to settle. The dust does not settle anymore. There is a new layer of it every quarter.

So the prescription is short. Pick the tool that came out this quarter and ship something real with it before the next one lands. Spend less time writing functions and more time writing specifications, permission boundaries, and the kind of review notes that actually change a pull request. Run an agent in auto-mode against a codebase you care about and feel where it fails, because that is where you still have to live. Stop optimising for being the person who can write the code. Optimise for being the person who knows which code is worth writing, and who can spot the moment it goes wrong.

The fourteen-thousand-stranger peer review on Stack Overflow is gone. The new peer review is you, on a Tuesday, reading a 600-line diff that an agent produced in eleven minutes, deciding whether to merge it.

If you want the full breakdown from Anthropic on how their own engineering team uses Claude, read it here. Fair warning, the productivity numbers will make you wonder what you actually did this morning.