JetBrains asked 15,000 developers how much code agents write. The answers add up to 112 percent

Add the three averages together and you get 112 percent of a month's output ^[16]. The survey put fully agent-generated, AI-assisted and fully manual code as three separate questions, answered in bands running 0 percent, 1-20 percent and upward to 100 percent ^[1][22]. Code a developer counts twice, once as agent output and again as assisted because they went back and edited it, lands in that 12 points of overshoot. So 47 percent is not a measured share of a repository. It is the midpoint of a self-reported band, and the bands do not tile.

The heavy end can be counted two ways, and JetBrains counted it both. Draw a line at 80 percent agent-generated code and 22 percent of respondents sit above it ^[5]. Cluster developers by behaviour instead and the agentic coder segment comes out at 31 percent, averaging 84 percent agent-written code ^[15]. That is a 9-point spread between two defensible readings of one dataset ^[19], which is roughly the size of most of the differences the release goes on to report.

Tool comparisons carry a selection problem that JetBrains names itself. Among developers whose most-used tool is Claude Code, 32 percent are above the 80 percent line; among Codex users it is 42 percent, and 37 percent of Codex users say they write no code at all without AI ^[7][8]. Cursor users average 58 percent agent-generated, 11 points above the overall figure ^[10][17]. JetBrains' own interpretation is that Codex has historically offered higher quotas and so draws advanced users looking for value, while Claude Code, at 39 percent adoption at work, has grown a mainstream audience ^[9][7]. On that reading the 10-point gap between Codex and Claude Code ^[18] is a fact about who quota pricing recruits, not about which agent writes more code.

The stack numbers behave more like a constraint than a preference. Go, JavaScript and TypeScript developers average 54 to 55 percent agent-generated code, seven to eight points above the overall average ^[11][20], while C and C++ developers still write 38 percent of theirs by hand ^[12]. Java and Python fall between ^[13]. Seniority runs against the usual assumption: about a quarter of senior developers are above the 80 percent line, ahead of juniors, who lean toward assistance rather than handing the work over ^[6].

The widest single split is geographic. Between 32 and 35 percent of developers in China, Japan and South Korea are above the 80 percent line, against about 16 percent in Europe and the UK ^[14]. So the fully agentic workflow is real, and it is roughly one respondent in five ^[5]. A plan that treats it as next quarter's baseline is extrapolating from the tail of a distribution whose own totals overrun by 12 points ^[16].