Computer-use agents move from benchmarks to real-world tasks

AgentConn analysis: top computer-use agents score 85% on OSWorld versus a ~72% human baseline, yet complete only about 20% of long-horizon workflows end to end. Meanwhile a Claude Code prompt that strips bloatware from Android TVs (CleanYourTV) drew nearly 7,000 GitHub stars in three days.

• Best computer-use agent scores 85% on OSWorld against a ~72% human baseline
• Agents complete only about 20% of long-horizon workflows end to end
• CleanYourTV hit 6,976 stars and 810 forks on GitHub in three days
• The prompt covers 92 Android TV and Google TV models from 31 vendors