Sharing my experience with **Razorpay's** agentic coding round for an **SDE-2** position.


This is a relatively new interview format. You get *90 minutes*, you own the design, and you implement using your own **AI coding agent** (bring your own keys/tooling). I used **Claude Code** with custom subagents and a personal conventions file. The round evaluates:


- your **low-level design**,

- how precisely you **prompt the agent**,

- your **process discipline** (incremental commits, reviewing each diff, testing as you go),

- and whether the final code **actually works**.



**Prep advice that helped me:**

- Get *fluent* with your harness **before** the round — it's not the place to learn shortcuts.

- Have your environment ready (toolchain, git, test runner) — the clock starts the moment the problem is shared.

- Build thoughtful, *general-purpose* tooling (a `CLAUDE.md`, test/review subagents) but **nothing pre-built for specific problem types** — that gets flagged.



Overall a fair test of whether you can *direct* AI rather than just lean on it. Will update with next rounds.


Sharing my experience with Razorpay's agentic coding round for an SDE-2 position.This is a relatively new interview format. You get 90 minutes, you own the design, and you implement using your own AI coding agent (bring your own keys/tooling). I used

Razorpay — Agentic Coding Round (SDE-2) experience