Benchmark
Next.js Agent Evals
Test coding agents on practical Next.js development tasks, with and without bundled-documentation guidance.
Explore the original evidence
Explore the original evaluation, its methods, and reported findings.
Open evalCoverage
Next.js application tasks evaluated across model and agent combinations. Current and superseded results are separate; changed tasks can retain older measurements until rerun.
Methodology
An agent modifies an app in an isolated sandbox, then withheld assertions check its work. The public leaderboard uses pass@4: a task passes if any of four attempts passes. Both conditions can access bundled documentation; AGENTS.md adds initial guidance. Infrastructure failures are rerun. Release automation can use a separate pass@1 protocol.