What it is
Office-centric GUI-agent benchmark with 100 long-horizon desktop tasks, executable evaluators, and released step-level demonstrations for demo-guided evaluation. 100 long-horizon office desktop tasks across 17 business tracks and 41 applications, plus 33 self-demonstrations with 212 subtasks and 3,989 captioned GUI steps.
Key results
- 100 test tasks across 17 business tracks and 41 apps
- 33 self-demonstrations with 212 subtasks and 3,989 steps
- Apache-2.0; Parquet tables plus mirrored runtime cache assets
Useful for benchmarking long-horizon computer-use agents on executable office workflows with or without demos.