TypeScript Dominates GitHub, Verification Becomes Central

The rise of AI coding agents shifts project constraints towards code review, altering developer evaluation and favoring typed languages. AI products are delivered via JavaScript interfaces, and GitHub is witnessing a significant increase in repositories related to LLM SDKs. GitHub's figures and corresponding surveys place TypeScript and the JavaScript family at the center of the landscape.

Verification Supplants Writing in the Delivery Cycle

In July 2026, a field report from OpenAI on coding agents used in scientific computing states that verification has become the limiting factor, rather than generation. This observation aligns with what many engineering teams have noted over the past year: interfaces that used to take weeks can now be assembled in an afternoon, and the slowest part becomes the evaluation of correctness, security, and maintainability. This shift changes expectations for JavaScript developers: typing speed, long used as a proxy for hiring competence, matters less than technical judgment in the face of generated code. Concrete pitfalls mark this review phase: a React effect can trigger twice during development; an unwary developer may spend a day tracking a false duplicate API call; a valid request in a testing environment can end up scanning an entire table in production; a misplaced authentication check may provide only an illusion of security, dissipated during actual testing. Generation capacity increases almost without limit with additional agents or subscriptions, while review capacity remains bounded by the number of engineers familiar with the system. Adding generators when review is saturated does not increase delivery throughput: the bottleneck simply shifts from writing to reviewing, and a team can double its generation in a week without increasing the number of available reviewers.

Recruitment and Evaluations Shift Towards Review Capacity

Most technical interviews continue to assess the ability to produce a functional solution, an aspect already heavily assisted by tools. Some companies are now testing the opposite skill by presenting candidates with generated code containing an error and measuring the time until correction. Recruitment agencies are also shifting in the same direction: Full Scale describes the JavaScript engineers it places based on their mastery of AI tools and product sense rather than the volume of lines written. Hiring a dedicated JavaScript developer is as much about increasing review capacity as it is about building capacity, in a context where an AI-generated connection flow can be shipped to production without human intervention.

Types Serve as Safeguards, and TypeScript Gains Ground

GitHub estimates that the appeal of typed languages lies in the security provided by their type systems in agent-assisted development. Generated code can run in dynamic languages and then fail later due to structural divergences that type checkers detect upstream. In 2026, professional usage of TypeScript is reported at 78%, up from 69% two years earlier; about 40% of developers code exclusively in TypeScript, compared to 6% in pure JavaScript. This trend is also explained by the nature of errors often overlooked in human review of 400 "reasonable" lines: function calls with missing fields, erroneous assumptions about the existence of a value, or handlers stuck on an outdated API response structure.

AI Products: Python for R&D, JavaScript for Interface, +178% in LLM Repositories

While training, evaluation, and most research tools remain in Python, the client-visible layer is predominantly written in JavaScript and TypeScript. Interfaces disseminate tokens, trigger tools, add approval steps, and explain system actions, whether they involve OpenAI models, Anthropic, or self-hosted open-weight models. By the end of 2025, GitHub reports over 1.1 million public repositories using an LLM SDK, a 178% increase year-over-year, a dynamic attributed mainly to applications rather than model development. Each pilot crossing the demonstration requires the construction of a user-oriented component, for which JavaScript frameworks remain the standard tooling.

Indicators Place TypeScript and JavaScript at the Top, with Methodological Reservations

TypeScript took the lead among languages on GitHub in August 2025, a shift presented as the most significant in a decade and occurring during the fastest adoption phase of coding agents. By October 2025, GitHub had 2.64 million monthly TypeScript contributors, a 66% increase year-over-year, and over a million developers wrote their first TypeScript code in 2025. On the survey side, the 2025 Stack Overflow survey, with over 49,000 responses, indicates that 66% of respondents use JavaScript, a rank the language has held almost every year since 2011. These converging trends come with reservations: GitHub's counting only measures activity on its platform, with a potential conflict of interest, and fashion effects influence public repositories. Nevertheless, the consistency between metrics and external surveys is deemed plausible given the scale of the movement, and the JavaScript family appears to be both the most used and the fastest growing on GitHub.

Training Data: Advantage to React and Concentration Effects

Models learn from public code, largely dominated by JavaScript, TypeScript, and projects around React. This results in more complete outputs for a typed React component—which compiles and often aligns with base standards—than for Svelte, Solid, or less widespread backend frameworks, where agents invent more APIs and require more corrections. Teams thus prioritize stacks most compatible with their tools, with productivity gaps accumulating over long cycles, and published outputs subsequently feeding into training, further reinforcing the advantage. This does not render a technical verdict: Solid and Svelte are considered high quality, and modern frameworks can surpass React in speed; the market has primarily favored what the models already knew. This training effect makes the adoption of new technologies more costly: a recent framework lacks a corpus, its agents struggle, teams avoid it, and no corpus is formed; typical funding timelines are insufficient to break this cycle, giving an advantage to frameworks that reached their milestones before 2023. At the enterprise level, betting product, tools, and recruitment on a single family of languages amounts to betting three times on the same thing: comfortable if dominance persists, costly in case of decline. Initial forecasts saw AI making language choice irrelevant; they are partially validated regarding the cost of writing in an unknown language, but the cost of understanding and ownership remains, and for many teams, it is this that guides the stack.