JetBrains: developers say agents now write 47% of their code

Professional developers now let AI agents fully write 47% of their code on average, according to a new JetBrains survey. The same survey finds 90% of developers use a coding agent at least weekly, and 68% use one daily. If you manage a codebase, that means agent-written code is no longer the exception in the pull requests you review. It is close to half of them.

JetBrains, the company behind IntelliJ and PyCharm, ran the survey from May to July 2026. Heise reports it covered 15,509 developers across eight languages. The findings come in two blog posts: one on which tools developers use, and one on how much code they let those tools write.

Three kinds of developer

The 47% average hides a split. JetBrains sorts respondents into three groups by how they work.

Group	Share of developers	How their code gets written
Agentic coders	31%	84% generated by agents
AI-assisted coders	47%	40% generated by agents
Manual coders	23%	75% written by hand

So nearly a third of developers already hand most of their typing to an agent. Another 47% mix agents and manual work. Under a quarter still write most of their code themselves. The middle group is the largest, and the report treats the top group as where the industry is heading.

Who has moved fastest

Seniority matters, and not the way many people expect. JetBrains reports that senior developers adopt agent-heavy workflows faster than juniors. About 25% of seniors generate 80% or more of their code with agents, a bigger share than among junior developers.

Region matters too. Developers in China, Japan, and South Korea lead adoption. JetBrains reports 32% to 35% of them generate 80% or more of their code with agents, against 16% in Europe and the UK.

Language shapes the numbers as well. Go, JavaScript, and TypeScript show the highest agent-written share, at 54% to 55% on average. C and C++ sit at the other end, where hand-written code still averages 38%.

Which tools developers actually use

Claude Code, Anthropic's command-line coding agent, is the most used tool. JetBrains reports 39% of professional developers use it at work, and 47% in the US. Heise notes that is up 21 percentage points since January 2026.

It is also sticky. JetBrains says Claude Code is the primary tool for 31% of developers, "which signifies an almost 80% conversion rate" from regular use to main tool. GitHub Copilot went the other way. Its adoption fell from 29% to 21% in a year, even though 79% of developers know it exists.

OpenAI's Codex grew fastest. JetBrains reports awareness jumped from 27% in January 2026 to 65% by mid-year, and adoption from 3% to 16%. Cursor slipped from 18% to 12%. The tool a developer picks also predicts how much they delegate. Among Codex users, 42% generate 80% or more of their code with the agent. Among Claude Code users, it is 32%.

What this means for developers

Review capacity is now the bottleneck, not writing capacity. If 47% of incoming code is agent-written, your review process was designed for a different ratio. Count how many lines a reviewer actually reads per week. If agents doubled the inflow and reviewers stayed flat, you are approving code nobody has read.

Do not read the seniority finding as "juniors are behind." Read it as seniors having the judgment to check agent output quickly, so delegation costs them less. The practical move for a junior is to build that checking skill on purpose. Read every agent diff until you can predict where it will be wrong.

The language gap is a hint about where agents are safe. Go, JavaScript, and TypeScript have large public corpora and fast test loops, and agents write over half the code there. C and C++ do not, and developers still write 38% by hand. If your codebase is in the second group, expect more manual correction, and budget for it.

Treat the tool-share chart as a hiring signal. Claude Code at 39% and Codex growing from 3% to 16% means new hires will arrive fluent in those two. A team standardised on a tool that is shrinking, like Copilot at 21%, should decide now whether that is a deliberate choice or an accident of history.

One number to hold loosely. The 47% figure is self-reported, and developers are not great at estimating what share of their own code an agent produced. It is the best industry-wide figure available. It is not a measurement of your repo. Run git blame against your own agent commits before you plan around it.