Survey finds 47% of professional developers' code now fully AI-agent-generated

The JetBrains survey quantifies a rapid transition in software development, where AI agents are no longer supplementary tools but primary code generators. The 47% average for fully agent-generated code suggests that for many developers, the role of AI has shifted from assistance to delegation. This change is not uniform, adoption varies by region, language, and experience level, indicating that workflows are evolving differently across the industry. Engineers must now decide how much of their codebase to trust to agents, a decision that will impact not just productivity but also long-term maintainability and debugging practices.

The data reveals stark differences in adoption by programming language. Developers working with Go, JavaScript, and TypeScript report the highest rates of agent-generated code, likely due to the prevalence of these languages in modern, high-velocity development environments. In contrast, C and C++ developers remain the least reliant on agents, possibly due to the precision and low-level control required in these languages. This split suggests that AI agents are currently better suited to certain types of development, and engineers in lower-adoption languages may face a steeper learning curve if they choose to integrate agents into their workflows.

Regional disparities in adoption are equally notable. Developers in East Asia, particularly China, Japan, and South Korea, are leading the shift toward agentic coding, with 32-35% generating over 80% of their code via agents. This is nearly double the rate seen in Europe and the UK. The reasons for this divide are unclear from the data, but it may reflect differences in tooling availability, cultural attitudes toward automation, or industry-specific demands. For global teams, these regional differences could create challenges in standardizing development practices or evaluating contributions from engineers with varying levels of agent reliance.

The survey also identifies distinct developer profiles based on AI usage. 'Agentic coders', who generate 84% of their code via agents, represent 31% of developers, while 'AI-assisted coders' make up 47%. This segmentation highlights that the transition to agentic development is not binary but exists on a spectrum. Engineers must consider where they fall on this spectrum and how it aligns with their project requirements, team dynamics, and personal skill development. Over-reliance on agents could lead to skill erosion, while underutilization might result in missed productivity gains.

The findings raise questions about the long-term implications of agentic coding. While the productivity benefits are clear, the survey does not address potential downsides, such as code quality, security vulnerabilities, or the maintainability of agent-generated code. Engineers will need to establish new best practices for reviewing, testing, and integrating agent-generated code into their projects. Additionally, the data suggests that tooling choices, such as Claude Code or Codex, influence adoption rates, which may drive further fragmentation in the AI-assisted development ecosystem.