Definition
Plain language
A huge benchmark of realistic professional computer-use tasks across hundreds of pieces of software.
As stated in the literature
A computer-use agent benchmark spanning over two hundred applications and twelve thousand tasks, grounded in GDP-weighted occupational data, with checklist-based verification.
Why it matters: It pushes computer-use evaluation beyond toy web pages toward the messy, varied desktop software actual knowledge workers depend on.
For example, an agent is graded on tasks like 'extract these line items from the PDF and enter them into the QuickBooks invoice form' across over 200 real applications.
Heard on the show
“CUA-World — that's what the authors call their benchmark — has over two hundred software and just over twelve thousand tasks.”Episode 017 — When the Agent Grades Its Own Homework: A Brutal New Benchmark for AI Workers