Thanks, I think I get it now. In our case I've dealt with this problem by refactoring the monolithic agent into smaller agents, smaller, more specific prompts, fewer, more relevant tools.

We've found that monolithic agents just don't perform that well.