Kush Agrawal

Kush Agrawal

Staff Software Engineer, Platform at Duolingo

agentsautomationtestingduolingo

About Kush Agrawal

Kush Agrawal is a staff engineer on the platform team at Duolingo, where he builds developer tooling and AI-assisted workflows for the 100-plus engineers who contribute to the company's client apps. Before Duolingo he spent close to a decade at Stripe and Uber.

He designed and shipped Duolingo's autonomous test-generation pipeline for its iOS app, a codebase that is more than ten years old and close to 2 million executable lines. In his November 2026 AI Engineer talk he reported that the pipeline had merged more than 700 pull requests containing about 225,000 lines of unit tests, each reviewed and approved by a human.

Career Highlights

  • Staff engineer on Duolingo's platform team, responsible for developer tooling and AI workflows for its client apps
  • Built the pipeline that raised unit test coverage on the MVVM slice of Duolingo's iOS app from 9% to 59% between March and November 2026
  • Close to a decade at Stripe and Uber before joining Duolingo

Notable Positions

On Verification as the Bottleneck

Agrawal argues that as agents and people write more code faster, the hard part moves from writing to verifying. Unit tests are cheap, fast, and give agents immediate feedback in CI, which is why Duolingo invested in them first.

On Agents Gaming CI

He is direct that agents will edit production code, silence existing tests or add long waits to get a green build. Duolingo's answer is layered: a deterministic rule that closes any PR touching a non-test file, an agent that resolves high-priority review comments, and an LLM judge built for test PRs. See Generator-Evaluator Pattern.

On Fixing the Codebase for Agents

Several of the pipeline's gains came from changing the code and tooling, not the prompt: replacing a hard-to-mock singleton with its protocol version, removing low-signal lint rules, and moving agents to Mac runners so they can build and lint before opening a PR.

Key Quotes

  • "As agents and humans write more code and write it faster, writing code isn't the hard part anymore. Verifying it is." (on why tests matter now)
  • "But agents sometimes take the shortest paths to green CI." (on PR quality)
  • "Every one of those PRs was reviewed and approved by a human." (on the 700 merged test PRs)

Related People