Jobs that involve computer programming are a small sector of the modern economy, but an influential one. The past couple of years have seen them changed dramatically by the introduction of AI systems that can assist with—and automate—significant amounts of coding work. (View Highlight)
In our previous Economic Index research, we found very disproportionate use of Claude by US workers in computer-related occupations: that is, there were many more conversations with Claude about computer-related tasks than one would predict from the number of people working in relevant jobs. It’s the same in the educational context: Computer Science degrees—which involve large amounts of coding—show highly disproportionate AI use. (View Highlight)
To understand these changes in more detail, we conducted an analysis of 500,000 coding-related interactions across Claude.ai (the “default” way that most people interact with Claude) and Claude Code (our new specialist coding “agent” that can independently accomplish chains of complex tasks using a variety of digital tools). (View Highlight)
The coding agent is used for more automation. 79% of conversations on Claude Code were identified as “automation”—where AI directly performs tasks—rather than “augmentation,” where AI collaborates with and enhances human capabilities (21%). In contrast, only 49% of Claude.ai conversations were classified as automation. This might imply that as AI agents become more commonplace, and as more agentic AI products are built, we should expect more automation of tasks. (View Highlight)
Coders commonly use AI to build user-facing apps. Web-development languages such as JavaScript and HTML were the most common programming languages used in our dataset, and user interface and user experience tasks were among the top coding uses. This suggests that jobs that center on making simple applications and user interfaces may face disruption from AI systems sooner than those focused purely on backend work. (View Highlight)
Startups are the main early adopters of Claude Code, while enterprises lag behind. In a preliminary analysis, we estimated that 33% of conversations on Claude Code served startup-related work, compared to only 13% identified as enterprise-relevant applications. The adoption gap suggests a divide between nimbler organizations using cutting-edge AI tools, and traditional enterprises. (View Highlight)
We also split automation and augmentation into several subtypes (as discussed in our previous work). “Feedback Loop” patterns, where Claude completes tasks autonomously but with help of human validation (for example, where the user sends any errors back to Claude), were nearly twice as common on Claude Code (35.8% of interactions) as Claude.ai (21.3%). “Directive” conversations, where Claude completed a task with minimal user interaction, were also higher on Claude Code (43.8%, versus 27.5% on Claude.ai). All the patterns of augmentation—including “Learning,” where the user acquires knowledge from the AI model—were substantially lower on Claude Code than on Claude.ai. (View Highlight)
These results illustrate the differences between specialist, coding-focused agents (in this case, Claude Code) and the more “standard” way that users interact with large language models (i.e., through a chatbot interface like Claude.ai). As more agentic products are released, we might see differences in the way AI is integrated into people’s jobs. At least in the case of coding, this might involve more automation of tasks. (View Highlight)
Overall, we found that developers commonly use Claude for building user interfaces and interactive elements for websites and mobile applications. Although no single language dominated, the primarily web-focused development languages of JavaScript and TypeScript together accounted for 31% of all queries, and HTML2 and CSS (other languages for user-facing code) together added another 28%. (View Highlight)
Back-end development languages (used for behind-the-scenes logic, databases, and infrastructure, as well as API and AI development) were also represented: notably, Python was at 14% of queries. However, Python serves dual purposes—both for back-end development and data analysis. Combined with SQL (another data-focused language, making up 6% of queries), these languages likely included many data science and analytics applications beyond traditional back-end development. (View Highlight)
These patterns further extend to the types of common coding tasks involving Claude. Two of the top five tasks were focused on user-facing app development: “UI/UX Component Development” and “Web & Mobile App Development” each accounted for 12% and 8% of conversations, respectively. Such tasks increasingly lend themselves to a phenomenon known as “vibe coding”—where developers of varying levels of experience describe their desired outcomes in natural language and let AI take the wheel on implementation details. (View Highlight)
We also analyzed which groups of developers might be using Claude. We used our analysis system to identify the type of project (e.g. a personal project vs. a project done for a startup) that best described users’ coding-related interactions. Because we don’t know the real-world context in which Claude’s responses were being used, these analyses rely on uncertain inferences from incomplete data. We therefore treat these findings as more preliminary than the ones described above. (View Highlight)
Startups appear to be the primary early adopters of Claude Code, and enterprise adoption lags behind. Startup work accounted for 32.9% of Claude Code conversations (nearly 20% higher than their Claude.ai usage), whereas enterprise work represented only 23.8% of Claude Code conversations (slightly below their 25.9% share on Claude.ai3). (View Highlight)
These adoption patterns mirror past technology shifts, where startups use new tools for competitive advantage while established organizations move more cautiously and often have detailed security checks in place before adopting new tools company-wide. AI’s general-purpose nature could accelerate this dynamic: If AI agents provide significant productivity gains, the gap between early and late adopters could translate into substantial competitive advantages. (View Highlight)
AI is fundamentally changing the ways developers work. Our analysis implies that this is particularly true where specialist agentic systems like Claude Code are used, is particularly strong for user-facing app development work, and might be giving particular advantages to startups as opposed to more established business enterprises. (View Highlight)
Our findings raise many questions. Will the prevalence of “feedback loops,” where humans are still involved in the process, persist as AI capabilities advance, or will we see a shift toward more complete automation? As AI systems become capable of building larger-scale pieces of software, will developers shift to mostly managing and guiding these systems, rather than writing code themselves? Which software development roles will change the most, and which might disappear entirely? (View Highlight)
The increasing coding skills of AI might also be especially consequential for AI development itself. Since so much of AI research and development relies on software, it’s possible that advancements in AI-assisted coding help to speed up breakthroughs, creating a positively-reinforcing cycle that accelerates AI progress even further. (View Highlight)