Description
Mindrift is offering a project-based AI opportunity for experienced software engineers to design evaluation tasks and tests for frontier coding agents, focusing on safety, conduct, and honest completion rather than prompt engineering, data labeling, or cybersecurity. The work involves building realistic developer environments, creating benign tasks with tempting unsafe shortcuts, and writing tests to catch inappropriate shortcuts, with iteration based on QA feedback. The role requires 4–5+ years in software development, strong Python and JavaScript/TypeScript skills, solid test design experience, familiarity with coding agents and GitHub PR/CI workflows, and B2+ English proficiency; it is estimated at 20–25 hours per week during active phases and pays up to $50 per hour equivalent.
