The main reason we've seen is a bad first impression: the agent misunderstands the initial prompt, doesn't load the available tools or starts doing something clearly wrong, and the user ctrl+c's instead of correcting it since it's faster to just start a new one.

The learning is that it is fixable. Better CLAUDE md instructions, clearer initial prompts, and skill configurations that reduce the uncertainty cut abandonment significantly in our team.