Next.js CRUD App Benchmark

Each agent was given the same prompt: scaffold a Next.js 14 app with Prisma, SQLite, NextAuth credentials-based auth, and full CRUD for a Task model. The app had to list, create, edit, and delete tasks, and protect all task routes behind authentication.

Criterion	Weight	Max points
Runs without errors	30%	30
Auth works end-to-end	25%	25
All CRUD operations work	25%	25
Code quality / structure	10%	10
Time to first working build	10%	10

• Same starter repo template for every agent.
• Same prompt, no follow-up help unless the agent explicitly asked a clarifying question.
• Time measured from first prompt to npm run build succeeding.
• Each run used Claude 4 Sonnet where the agent supported model selection.

• Claude Code won on integration speed. Its diff-aware editing and test-running loop caught missing environment variables early.
• Aider was the most cost-effective. Slower than Claude Code but produced the cleanest git history and least rework.
• Hermes Agent needed more steering. Strong on structure, weaker on nuanced framework conventions.
• OpenHands spent a lot of tokens exploring. It generated many files but failed to connect auth to the task routes.

• We tested one representative task, not every Next.js pattern.
• Cost figures are approximate and include retries from the agent's own loop.
• Your mileage will vary based on how well your existing project matches the agent's training distribution.

• Claude Code: complex greenfield apps where speed and correctness matter more than cost.
• Aider: existing codebases where you want surgical, reviewable edits.
• Hermes Agent: workflows where messaging, memory, or multi-step skills are the real requirement.
• OpenHands: research and experimentation; not yet our first choice for production scaffolding.