Stanford AI Index 2026: AI Agents Jump from 12% to 66% on Computer Tasks but 89% Never Reach Production

Stanford HAI·medium signal

The Stanford HAI 2026 AI Index Report shows agent accuracy on OSWorld computer tasks rose from 12% to 66.3% (within 6 points of human performance), Terminal-Bench from 20% to 77.3%, and cybersecurity agents from 15% to 93%. However, the report also reveals 89% of deployed AI agents never reach production, and the best agents still perform only half as well as PhD-level experts on complex scientific tasks, creating what Stanford calls a 'jagged frontier' where agents that solve 66% of computer tasks still can't reliably read analog clocks.