1. They require a little more precise notions of what needs to be done by the system (like, go here, click here, then do this, then do that etc.)
2. To perform complex tasks, the few tools that exist, they need some knowledge of the system itself (like telling it to call a specific function or a class etc.)
My goal with this project (I call it `pyba`) was to abstract everything from the user's side. I have two "exploratory" modes built into it, these are for Depth First Search (wherein, it will pick one train of thought, and execute that in full, then revert), or Breadth First Search (where it executes multiple tangential plans in parallel).
All an analyst has to do is type out in words, everything they already know and can share, and what all they'd like to find out more. To ensure that there are no roadblocks in this journey, I also have hardcoded logins, which read your credentials from the environment and log you in to websites like gmail, facebook and instagram (so your credentials never go to the LLM).
And of course, you can use it for normal stuff as well (like tell it to scroll reels for you if you want)!