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.
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.