DesktopBench
RL environments for professional desktop work.
Can agents actually do complex work inside professional software?
Economically valuable work happens inside professional tools.
Frontier agents can handle short edits, but still break on long-horizon professional software workflows because they lack the RL data and environments needed to practice real work.
So we build resettable, model-native software environments where agents can practice, fail, and improve. Read our manifesto.
Benchmark
DesktopBench
Public scores on professional desktop workflows.
Compare agents on resettable tasks with programmatic verifiers and trace-backed results.
Training
RL Environments
Train on resettable desktop tasks.
Use environments with deterministic resets, rewards, verifiers, and trajectory traces.
Suites
Workflows
Browse real work domains.
Commerce, accounting, support, sales, and office work built into desktop task suites.