Rapid web app development with Devin – A Developer's Perspective (blog.scottlogic.com)

Cognition Labs’ Devin — an “agentic” AI software engineer that plans, executes and tests code with minimal human direction — was put to the test converting a shelved carbon-emissions Excel workbook into a production-ready TypeScript/React web app. Over about seven days the author and Devin implemented the original calculators plus two new ones, a dashboard, import/export (including TCS schema), unit tests and Playwright end-to-end tests. Devin links to a repo, spins up a VM, creates branches, raises PRs, and provides plans with confidence ratings; the exploratory build consumed ≈155 ACUs (~$350). The experiment highlights both opportunity and caveats for the AI/ML community: agentic systems can massively accelerate the software lifecycle and produce complete features and tests, but they require experienced engineers to guide architecture, validate core logic, control bloat and manage technical debt. Devin works best on live, inspectable codebases (struggled with Electron), benefits from iterative prompts and context-management (start new sessions to avoid rot), and changes developer workflows (parallel sessions, “fire-and-forget” for low-risk tasks). Cost transparency, verification at scale, and automated refactoring remain crucial — Devin can generate volume quickly, but human oversight is still essential to ensure secure, maintainable production code.