Vibe coding features in a complex SaaS app (reflag.com)

An engineer used Cursor in “Agent mode” to implement a complete new feature in Reflag—a React/Node.js SaaS with a TypeScript-shared REST API and Prisma-managed DB—adding an “owner” property to feature flags and a “my flags” view. The agent scanned the repo, updated schema.prisma, Zod validators and shared types, generated UI components, created a feature flag through Reflag’s MCP, and ran the CLI to refresh local types. The AGENTS.md file with repo rules proved essential for smooth automation. Iterative prompting fixed several issues: the agent initially picked the wrong UI component, caused runtime/validation errors (missing name/orid on a Select), and implemented the view differently than intended; human-directed rounds produced tests and a critical security check preventing assignment of users outside the organization. This experiment shows agents can produce meaningful, end-to-end changes in structured codebases but require human oversight and tight prompts. Key takeaways for the AI/ML community: agent success depends on existing patterns and clear repo guidance (AGENTS.md); type-safety workflows (CLI/type updates) can be automated by agents; and security, tests, and UX details still need humans in the loop. Expect more agent-generated PRs—and consequently more code review, merge conflict management, and scrutiny around subtle bugs or security gaps in agent-authored code.