OSWorld 2.0 leaves agents at 20.6%

OSWorld 2.0 puts computer-use agents through long-horizon work, and the strongest system in the benchmark completed only 20.6% of tasks outright, according to Snorkel AI’s report on the benchmark. The tasks span messy, multi-application workflows rather than short scripted clicks, which makes the result a useful counterweight to benchmark audits that find most evaluation questions redundant. Builders should file 20.6% as a boundary marker: keep computer-use agents supervised until they prove reliability on long, changing workflows.