I Broke a Browser Agent on Purpose. Spotting the Fault Was Not the Same as Fixing It

An experiment was conducted to analyze the failure modes of an AI browser agent by intentionally injecting faults. Correct decisions from a Claude Haiku 5.5 model were replayed through four types of broken browser actions, such as incorrect click coordinates due to screenshot resizing. The study found that simply detecting a fault with a check does not mean the agent can recover and complete its task. Different verification policies were tested, revealing that a simple page check missed half of the incorrect clicks and raised false alarms on actions within embedded frames.