๐ a16z: Can Agents Use a Computer Yet? We've Got the Data
a16z surveyed operators running computer-use agents in production and found the leap from demo to deployable happened fast.
โข Best models jumped from 42% to 85% task success on the OSWorld-Verified benchmark in a year โ above the ~72% human baseline.
โข Running an agent costs roughly $6โ8/hour, break-even with offshore BPO (~$10/hr) and a 70โ80% margin vs. US back-office labor (~$30โ45/hr).
โข The moat is shifting from the model to context โ verification, escalation, and company-specific workflow knowledge now separate winners from demos.
๐ Read more
#a16z