The results of the JetBrains Developer Ecosystem Survey 2026 indicate that the use of coding agents has moved beyond the limited experimentation stage, but has not yet become a complete working method for the majority of developers. More than 15,000 professional developers who participated in the survey between May and July 2026 reported the percentage of code they produced for work during the previous month that was either generated entirely by AI agents, written with AI assistance, or written manually without any assistance.
According to the participant averages published by JetBrains, around 47% of the code was generated entirely by agents, around 38% was written with AI assistance, and approximately 27% was written entirely manually. The study presents these percentages as they appeared in its results, although their total exceeds 100%, which is a point that warrants methodological review before treating them as a direct arithmetic measure of the distribution of mutually exclusive categories.
Full reliance remains a minority practice
Despite the decline in manual coding, most developers do not use agents to write all of their code. JetBrains says that more than half of developers manually write less than 20% of their code, and that one in five developers does not write any code without AI assistance. By contrast, those who generate more than 80% of their code with agents account for only around 22% of participants.
This result presents a different picture from common claims that “100% of code” has become automatically generated. The overall trend is clear, but actual use is closer to a combination of full generation, partial assistance, and manual work than to the wholesale replacement of developers.
More experienced developers lead adoption
The study shows that highly experienced developers are among the first groups to shift a broad share of code writing to agents. Around one-quarter of senior developers generate more than 80% of their code with agents, a higher proportion than that recorded among junior developers, who lean more toward AI-assisted work rather than full reliance on agents.
This does not mean that all experienced developers have adopted the agent-based model; JetBrains confirms that there is considerable variation within this group itself. In practice, the result suggests that experience does not necessarily lead developers to abandon manual coding, but it may make them more willing to delegate a larger part of the generation process.
Differences between tools and programming languages
Among Claude Code users, around 32% generate more than 80% of their code with agents. The proportion rises to 42% among Codex users, while the share of those who do not write any code without AI assistance reaches 37% in the Codex user group. Cursor users are relatively similar to Claude Code users; the average share of agent-generated code among them is 58%, and 28% use agents to generate more than 80% of their code.
JetBrains attributes these differences to the nature of the user bases, rather than to the overall quality of one tool. Claude Code users have become a broad segment, while the Codex user base may include a higher proportion of developers who are advanced in their use of agents, according to the study’s interpretation.
At the programming-language level, Go, JavaScript, and TypeScript developers record the highest average shares of agent-generated code, at 54% to 55%. By contrast, C and C++ developers retain the highest share of manually written code, averaging 38%. Java and Python developers fall in the middle, with averages ranging from 48% to 51%.
Geography reveals sharp differences
Regional differences are among the survey’s most prominent findings. In East Asia—particularly China, Japan, and South Korea—the share of developers who generate more than 80% of their code with agents is approximately 32% to 35%. This is close to twice the level recorded in Europe and the United Kingdom, where it is around 16%.
What is changing in practice?
JetBrains divided participants into three patterns. “Agentic developers” account for approximately 31% of developers and generate an average of 84% of their code entirely with agents. “AI-assisted developers” account for approximately 47%, with an average of 40% of code generated entirely and 60% written with AI assistance. “Manual developers” represent approximately 23% and write an average of 75% of their code manually, with limited use of full generation.
The most important reading of these figures is that the shift is not occurring as a binary transition between manual programming and agentic programming, but across a spectrum of working patterns. This spectrum is influenced by experience, tool, programming language, and region, meaning that measuring agent adoption with a single usage percentage may conceal major differences in actual working methods.
The limits of these findings remain tied to the fact that they are based on developers’ answers about their own estimates, while the extracted text does not provide sufficient detail about the methodology used to classify the patterns or how the averages were calculated. JetBrains has also announced that it will publish additional materials from the same survey on agentic development.