Design a Ticket-to-PR Coding Agent

A coding agent that takes an engineering ticket, finds the problem in the codebase, writes the fix, and opens a pull request is the agentic question: a generative system that acts instead of answering. Interviewers at infrastructure and developer-tool companies (Datadog asks it in both the phone screen and the onsite) use it to check whether we can reason about an LLM that runs tools over minutes, in a sandbox, against a codebase far larger than any context window, with a reviewer's time as the scarce resource. This walkthrough follows the Interview Framework, with the same announced deviation the RAG support assistant uses: evaluation is agreed on right after the objective, because "good" has to be defined before any architecture is worth drawing.

The common way to fall short on this question is to spend the whole session on infrastructure: the retrieval index, the tool protocol, the sandbox. Those matter, and they get their sections below, but they are the plumbing around a model. The LLM-side decisions (what enters the context, how the agent loop is bounded, how the output is verified, what it costs per ticket, and how the system learns from merge outcomes) carry equal weight, and a design that treats the model as a black box reads as incomplete.