▶ Watch ↗12:38
Conference affiliation: Duolingo
Kush Agrawal is a staff software engineer at Duolingo who designed and shipped its autonomous test-generation pipeline, carrying agent-written tests through CI, automated review, and human approval. By November 2026, the pipeline had merged more than 700 pull requests containing about 225,000 lines of unit tests into Duolingo’s 1.8-million-line iOS codebase. Kush builds tooling, infrastructure, and AI-assisted engineering standards supporting more than 100 engineers, with verification becoming a larger part of that work as agents produce more code.
Before Duolingo, Kush spent nearly a decade at Uber and Stripe. At Uber, he built client infrastructure that expanded the Rider app’s capabilities, including a server-driven home screen launched in more than 70 countries. His earlier mobile engineering work also included growth projects involving cash payments and PIN-based dispatch.
At Stripe, Kush helped build the Express app from the ground up; by November 2026, the app had more than 1.2 million downloads. He also shipped payment features in the Stripe Dashboard app that drove more than $1 billion in volume. He led the build of Global Payouts, which debuted at Stripe Sessions in 2025 and lets businesses onboard and pay recipients in more than 100 countries directly from their Stripe balance.
Kush also founded Hyre, which participated in Y Combinator, and IHeartCode, which participated in Mozilla WebFWD. He holds a B.S. in electrical engineering and computer sciences from UC Berkeley.
Kush’s writing on AI-assisted Swift development distinguishes knowledge of a programming language from understanding a particular application. He recommends defining requirements, asking the assistant to restate its understanding, and reviewing its proposed implementation before it generates code.
His skeleton files make architectural choices tangible. The engineer creates protocols, empty classes, and minimal views that establish where responsibilities belong; the assistant fills in the implementation. Kush describes using AI to add a SwiftUI file-watching panel to an older AppKit application, where generated code had to fit the application’s existing structure.
His approach to repository instructions addresses what an assistant needs to know for each task. Task-specific rules supply relevant guidance without loading every repository convention into every interaction. Maintaining them has a cost: different assistants need overlapping documentation, and engineers must keep it useful. Kush calls for evaluating the rules themselves—whether an instruction improves code acceptance, whether a revision makes results worse, and which guidance earns its maintenance cost.
Duolingo’s test-generation system applies this structured approach across a large codebase. One Temporal workflow selects untested code using coverage data and sends Claude Code agents to write tests and open pull requests. Another manages those changes, attempts to repair CI failures, runs automated review, and assigns a human reviewer. Engineers can also request tests for their own changes by adding a label to a pull request.
Kush designed and shipped the pipeline; his colleagues contributed supporting infrastructure and changes that made the application easier to test. Several decisions explain how the team made generated tests usable:
From March to November 2026, coverage of Duolingo’s core model-view-viewmodel code rose from 9% to 59%, while overall iOS coverage rose from 11% to 24%. The same identify, write, verify, and resolve loop also supports modularization, UIKit-to-SwiftUI rewrites, Android coroutine migrations, and snapshot test generation at Duolingo. Its reusable contribution is the process around code generation: selecting work, checking the result, resolving failures, and bringing a change to a human reviewer.
▶ Watch ↗12:38
Loading more talks and highlights…