/
Published on 03.03.2026
TLDR: A massive survey of nearly 1,000 software engineers shows Claude Code rocketing from zero to the number one AI coding tool in just eight months. 95% of respondents now use AI tools weekly, 55% regularly use AI agents, and Anthropic's models dominate coding tasks by a margin that is frankly not even close.
The Pragmatic Engineer just dropped their latest AI tooling survey results, and let me tell you, the data here is fascinating -- not because it confirms what many of us suspected, but because the speed and magnitude of the shifts are genuinely startling. This is one of those surveys that will be referenced for the next year as a benchmark for where the industry actually stands.
Let's start with the headline number: Claude Code, a terminal-first coding agent released in May 2025, is already the most-used AI coding tool among survey respondents. Eight months. That's it. It overtook GitHub Copilot, which had a multi-year head start, and Cursor, which has been growing aggressively. For perspective, Claude Code's current market penetration is roughly equivalent to where GitHub Copilot was when it was the undisputed leader three years ago. The AI tooling market moves at a pace that makes even seasoned technologists uncomfortable.
What makes this even more interesting is the "love" factor. When asked which tools they enjoy most, a staggering 46% of respondents named Claude Code. Cursor came in at 19%, and GitHub Copilot at just 9%. That gap between usage satisfaction and market presence is the kind of thing that should make product teams at competing companies lose sleep. Directors and senior engineering leaders are especially enthusiastic about Claude Code -- it is twice as popular with these folks compared to less senior levels. Meanwhile, Cursor actually gets less love as seniority increases.
Now, here is something the article largely glosses over: what exactly is it about the terminal-first paradigm that is winning? The survey hints at it -- the common setup is a split screen with Claude Code in a terminal driving work, and an IDE open to review changes. This is fundamentally different from the inline-completion model that GitHub Copilot popularized. It suggests engineers are shifting from "AI helps me type faster" to "AI does work while I supervise." That is a philosophical shift worth paying attention to.
The survey puts a definitive number on something we have all been sensing: 95% of respondents use AI tools at least weekly. Seventy-five percent use AI for at least half of their software engineering work. Fifty-six percent report doing 70% or more of their work with AI assistance. These are not early adopter numbers. This is the new baseline.
But here is where I want to push back a bit. The article presents these numbers without much critical examination of what "using AI" actually means. Is tabbing through autocomplete suggestions the same as delegating entire features to an agent? The survey does attempt to differentiate, noting that 55% now regularly use AI agents, but the conflation of "I use Copilot to complete my function signatures" with "I let Claude Code write my pull requests" risks overstating how deeply AI is truly integrated. The real story is probably more nuanced -- there is a bimodal distribution hiding in these averages, with some engineers going all-in and others using AI as a glorified snippet generator.
Fifty-five percent of respondents say they regularly use AI agents. That is 507 people out of the survey, and it is a massive jump from eighteen months ago when agents were basically experimental curiosities. The most common agent use cases are code review, automating tedious tasks, bug investigation, and what one respondent beautifully called "crafting" or "weaving" code together with the agent.
Staff-plus engineers are the heaviest agent users at 63.5%, ahead of regular engineers at 49.7% and engineering managers at 46.1%. This is a telling detail -- the most experienced engineers, the ones who theoretically have the best judgment about code quality, are the ones leaning hardest into agent-assisted workflows. That should quiet some of the skepticism about agents producing lower-quality output.
There is a fascinating correlation buried in this data: agent users are nearly twice as likely to feel excited about AI compared to non-users. Non-users are twice as likely to be skeptical. The article raises the right question -- is this correlation or causation? -- but does not really engage with the uncomfortable possibility that this could simply be a self-selection effect. People who are already enthusiastic about technology try new things. That does not mean agents would convert the skeptics. What is missing from this analysis is any data on engineers who tried agents, had a bad experience, and stopped. That churn number would be far more informative than the enthusiasm of current users.
One of the most practically relevant findings is how company size shapes tool choice. Claude Code is used by 75% of the smallest companies and teams. At companies with 10,000-plus employees, GitHub Copilot dominates at 56%. The article correctly identifies this as a procurement story rather than a preference story -- Microsoft's enterprise sales machine is doing what it does best.
But there is a deeper architectural implication here. Massive companies often build their own internal coding agents -- Block has Goose, Meta has its own, Google has Jetski. This fragmentation at the top end of the market suggests we might be heading toward a world where the largest companies treat AI coding tools as core infrastructure they need to own and customize, while smaller companies ride the wave of best-of-breed commercial tools. For architects and team leads at mid-size companies, this raises uncomfortable questions about build-versus-buy that will only get more pressing.
The article also notes that the feeling of being empowered to experiment with tools correlates strongly with Claude Code availability. At companies with more bureaucratic approval processes, engineers feel less able to experiment -- not just with Claude Code, but with any new tool. This is a warning shot for engineering leadership: your procurement process might be your biggest competitive disadvantage in AI adoption.
On the model side, Anthropic's Opus and Sonnet models dominate coding tasks by more mentions than all other models combined. This is not even close, as the article notes. One in eight respondents just use whatever model is the default, which raises an interesting point about how many engineers are unknowingly hamstrung by suboptimal default configurations.
The survey was conducted before Opus 4.6, Sonnet 4.6, and GPT-5.3 were released, so the landscape may have shifted. But the trajectory is clear -- for coding-specific tasks, Anthropic has established a substantial lead in developer preference. Whether that holds as competition intensifies is the trillion-dollar question.
Here is what the article avoids thinking about. First, there is zero data on code quality outcomes. Are teams shipping better software with AI? Faster, sure, but better? Second, there is no discussion of the security implications of 56% of engineers doing 70% or more of their work with AI. What does the vulnerability surface look like when the majority of code is AI-assisted? Third, the survey is self-selected Pragmatic Engineer readers -- a technically sophisticated, engaged audience. The broader industry, including the vast number of engineers at non-tech companies writing internal CRUD applications, almost certainly has very different adoption patterns.