Design an Agentic System for Closing Open Issues on GitHub
Design an LLM-powered system that automatically reviews a repository's open GitHub issues and decides which ones can be closed based on defined criteria, including analyzing issue content, comment threads, and linked pull requests.
Question Timeline
See when this question was last asked and where, including any notes left by other candidates.
Early October, 2026
Design an LLM-powered issue closer A company has a large open-source repository on GitHub. GitHub is its issue tracker, but the problem would be the same in Jira or a similar tool. The repo has built up a big backlog of open issues and enhancement requests. Many of them have already been resolved or implemented, but nobody closed them. Design a system that uses an LLM to review a repository's open issues and decide which ones can be closed based on some criteria. It can run inside GitHub (for example, as a CI job) or as a separate service. What the inputs look like: - Some issues have no content at all. - Some have long comment threads. - Some have one or more linked PRs. One issue may take several PRs to resolve. How the interviewers steered it: - The interviewer said this isn't a typical system design interview, so not caching and databases. The focus is on how you approach a business problem, edge cases, and requirements that change. - Detecting low-quality or spam PRs was ruled out of scope. - Partway through, they moved from overall architecture to a single issue in one specific repo. They asked for: - the criteria that would make an issue closeable - how you would test whether those criteria hold - a sample prompt, written out in full, that the process would run - The idea was to get one issue working first, then see whether the approach generalizes. Follow-up questions: 1. This runs as a service. How do you make the LLM's output structured enough for code to process it, with no human involved? 2. The point is to take an action, not only produce a report. How does the system actually close issues? 3. How do you see what the agent is doing, for example whether it actually read the issue and the linked PRs? Context: about 30 minutes, with a shared whiteboard.
Your account is free and you can post anonymously if you choose.