AITowards Data Scienceabout 2 hours ago
Researcher Tests Broken Browser Agent
14 min read
A researcher tested a deliberately broken browser agent against 40 layouts using Claude Haiku 5.5, finding that spotting faults differs from fixing them. The agent's score rose from 31.0 to 57.0 out of 100 with fixes. A 900 pixel wide image scaled to 1200 pixels caused clicks to land 300 pixels off target.
Level
Hype check
Tap to vote and see what everyone thinks.
