@addyosmani: Make Your Agent Prove Its Work

Evidence, not a claim, that the work is done.

“Want better quality output from your agent? Give it a way to verify its own work,” @addyosmani writes. On top of end-to-end tests, Claude gives him recorded videos and screenshots as part of the evidence that work is done.

He points to Warp. There, @varoonkodithala explains, every agent in Warp Factories gets a computer to open and click through the apps it built, take screenshots and recordings, and post them for human review. Three to four minutes of activity are edited down to 15 clips. He says it cut their time to merge.

Cursor says its agents can now control your computer, driving desktop apps like the iOS Simulator to test the features they build. @poteto reposted it.

Osmani’s bet: verification becomes one of the most important problems to work on this year.