Tomik99 survey finds 95% of devs use AI
- The Pragmatic Engineer’s March 2026 AI tooling survey of 906 engineers put AI use near saturation, with 95% using tools weekly or more. - The sharpest shift was agents: 55% said they use them regularly, and staff-plus engineers were the heaviest adopters for architecture and automation. - This matters because AI coding has moved from autocomplete to delegated work — but verification, rollout, and ownership are now the bottlenecks.
Software engineering has crossed a line with AI. This is not about a few people tab-completing boilerplate anymore. In The Pragmatic Engineer’s March 2026 survey, 95% of 906 engineers and engineering leaders said they use AI tools at least weekly, and 55% said they regularly use agents — tools that do chunks of work with more autonomy than a chat box. The interesting part is not that AI is common. It’s that the job is starting to split between people who use AI as a faster keyboard and people who use it as a junior teammate. (aiproductivity.ai) ### What kind of survey was this? This was not a mass-market poll of random coders. The sample was 906 software engineers and engineering leaders, with a median experience band of 11 to 15 years, and a respondent base concentrated in Europe and the US. That matters because the numbers are closer to “how experienced working engineers are using AI now” than(aiproductivity.ai)ace usage over casual experimentation. (aiproductivity.ai) ### Why is 95% the wrong headline? Because “uses AI weekly” sounds dramatic, but it hides the more important shift. Weekly use can mean asking for regex help, writing tests, summarizing docs, or generating a migration script once in a while. The bigger signal is that 75% of respondents said AI helps with more than half their work, and 70% said they use tw(aiproductivity.ai)DE plugin. It’s a stack. (neuralcoretech.com) ### What changed with agents? Agents went from fringe to normal fast. The survey says 55% of respondents now use them regularly, up from near-zero about 18 months earlier. That is a very different pattern from plain autocomplete. An agent can inspect a codebase, propose a plan, edit files, run commands, and loop on errors. In practice, that makes it useful for larger u(neuralcoretech.com) not just line-by-line suggestions. (aiproductivity.ai) ### Who is actually using agents hardest? Senior people. Staff-plus engineers showed the heaviest usage, especially for architecture work and automation. Junior engineers still use AI heavily, but more for code completion, tests, and getting unstuck. That split makes sense. The more context a task needs — system design, migration strategy, orchestration a(aiproductivity.ai)tor lacks judgment. AI is amplifying leverage, but mostly where leverage already existed. (newsletter.pragmaticengineer.com) ### Which tools are winning? Claude Code was the standout. The survey coverage around the release says it went from not existing to the most-used or most-loved coding tool in about eight months after its May 2025 launch. In the “most loved” ranking, it led with 46%, ahead of Cursor at 19% and GitHub Copilot at 9%. That does not mean Copilot vanished — lar(newsletter.pragmaticengineer.com)veloper preference has clearly shifted toward more agentic, terminal-first tools. (aiproductivity.ai) ### So what is the real bottleneck now? Verification. The more autonomous the tool, the less the problem is typing speed and the more the problem is trust. Another 2025 developer survey showed experienced developers are the least likely to highly trust AI output and the most likely to highly distrust it. That lines up with what teams keep running into: ag(aiproductivity.ai)the edge cases, and decide whether the output belongs in production. (survey.stackoverflow.co) ### Why does this matter for teams? Because rollout is becoming an engineering design problem, not a software-buying problem. The companies getting value are not just handing everyone a license and hoping for magic. They are deciding where agents can act, what tools they can call, how outputs get tested, and who is accountable when the model is wrong. That is why the survey’s senio(survey.stackoverflow.co)ow design and governance, not just individual productivity. (newsletter.pragmaticengineer.com) ### Bottom line? The survey’s 95% number tells you AI is everywhere. The 55% agent number tells you the shape of software work is changing. Coding assistants helped engineers write faster. Agents are starting to let experienced engineers delegate. That is a much bigger shift — and it puts the value on judgment, review, and system ownership, not just raw implementation speed. (aiproductivity.ai)
