For years, a developer’s job consisted mainly of writing code.
Today, some explain that their role is starting to change.
They write less.
They supervise more.
The “Claudeholics” phenomenon
In an article published in late May 2026, WIRED describes the appearance of a new category of developers nicknamed the “Claudeholics”.
The term comes from users who have become extremely dependent on Claude Code.
Some explain that they spend most of their day:
- delegating;
- correcting;
- supervising;
- orchestrating several AI agents.
More than coding themselves.
Agents that create other agents
The latest versions of Claude Code can:
- run code;
- modify files;
- use tools;
- launch sub-agents;
- handle complex tasks over several hours.
Anthropic even states that its Opus 4.5 model scored higher than every human candidate who took its internal engineering recruitment test.
That capability is not an isolated claim either: it sits on the same curve we described the day before, when a study of MCP tools showed that the tools agents use are increasingly tools that act rather than tools that read.
A new way of developing
The change observed is less technical than cultural.
Many developers now explain that they work more like:
- supervisors;
- coordinators;
- validators;
than like traditional programmers.
The AI produces.
The human checks.
Which raises the obvious follow-up: produces how well, exactly? That is a measurable question rather than a feeling, and we put the two leading models through it side by side in our comparison of the best AI for coding.
A study already shows a measurable impact
Research published on 25 May 2026 covering more than 5,800 developers indicates that adopting Claude Code is associated with:
- more commits;
- more projects;
- more languages used;
- more technical experimentation.
The researchers remain cautious about causality.
But the figures already show a real change in behaviour.
What to remember
Claude Code is no longer only a coding assistant.
For some developers, it is gradually becoming a complete production layer.
How complete is precisely what we tried to find out by leaving two agents to work on their own on a Windows machine: what each of them finishes without help, and where each one stops. The part that stops is the part a human still supervises.
And a question is starting to appear:
will tomorrow’s developer still mainly write code, or mostly supervise AI agents?