GitHub
S03
HardPremiumDesign an Autonomous Coding (SWE) Agent
Design an agent that takes a GitHub issue, finds the fix in a large codebase, runs the tests in a sandbox and opens a reviewable pull request.
Code AgentsSandboxingTest-DrivenRepo NavigationEvaluation
Key Requirements
- Navigate a large, unfamiliar codebase without loading it into context
- Run every task in its own sandbox and branch, with no credentials inside
- Let tests decide done, and stop the agent from editing them to get there
- Survive crashes and long runs without duplicate pushes or PRs
- Produce small, explained pull requests a person can review quickly
Interviewer Follow-ups
- Q1How do you prevent the agent from introducing security vulnerabilities?
- Q2How does it handle flaky tests vs. real failures?
- Q3How do you find the relevant files instead of loading the whole repo?
- Q4When would you let it open or merge PRs without a human?
Loading...