AI coding agents were supposed to erase language choice. They sharpened it.

Two years into widespread AI coding tool adoption, the 2023 prediction that coding agents would make language choice irrelevant has failed for developers, hiring managers, and the CTOs who sold stack-agnostic engineering orgs. TypeScript took the top language slot on GitHub in August 2025, and the Octoverse 2025 report counted 2.64 million monthly TypeScript contributors, a 66% year-over-year jump. GitHub called it the largest shift in its language rankings in a decade, the same period in which AI coding tools moved from preview to default at most companies. The data reads less like a TypeScript victory lap than a snapshot of where the training corpus already lives.

The 2023 stack-agnostic pitch held that agents would abstract over syntax, the corpus would carry the rest, and teams could hire for problem-solving and pick stacks later. The 2025 numbers point the other way. JavaScript still commands 66% of the Stack Overflow 2025 developer survey on 49,000-respondent data, the same share it has held since 2011 (secondary source). Over a million developers wrote their first TypeScript code on GitHub during 2025, on top of an already dominant base. The JavaScript family did not just hold its ground; it absorbed the new wave of agent-era developers.

The mechanism is a feedback loop, visible in the gap between two prompts. A typed React component for a standard CRUD form returns clean, idiomatic output. A Svelte or Solid request, or a backend task in anything less popular, returns uneven output, sometimes within a day. The reason is upstream: coding models train on what is published, published code is concentrated in JavaScript, TypeScript, and React, so the model is well-conditioned for those stacks and thin on everything else. The original analysis puts the contrast in one sentence, and InfoWorld's TypeScript coverage tracks the same ranking shift from the trade press side. The Octoverse landing page calls the period the year AI became a default collaborator.

A second-order effect shows up in how the major labs approach training data. Coding-agent teams have an incentive to keep training on the published web as it stands, because the published web is what their customers write in, and their customers overwhelmingly write in JavaScript and TypeScript. Curating a more balanced corpus would lift long-tail quality and lower JavaScript-family quality in the same step. No major lab has announced a deliberate rebalancing effort, and the year-over-year corpus snapshot shows the JavaScript family's share of new code growing, not shrinking.

The same data is not the only data. GitHub's count is GitHub-only activity, and GitHub publishes the report; the analysis piece itself flags the conflict of interest. Fashion cycles shape what developers publish, so the corpus is concentrated by taste and reach together. The Stack Overflow survey is the cross-check, and the two converge on the JavaScript family, but neither can settle whether agent adoption caused the concentration or rode it. The safe read is correlation with a plausible mechanism, not a closed cause.

Developers picking frameworks in 2026 now face a cost gradient that did not exist in 2023. The agent's first answer is much better on a JavaScript-family stack than on a long-tail alternative, and the second- and third-pass iterations are where most of the project time lives. Hiring managers writing job posts cannot treat language as fungible, because the agent works differently across stacks and the team's accumulated context is the asset. CTOs who pitched language-agnostic orgs in 2023 have to decide how much of that promise still holds. A future agent trained on a more balanced corpus would invert the mechanism. The question for the next two years is whether any lab has the incentive to build it, and the team that ships it would just move the menu again, because the training set is the menu.