I am amazed by the jump in usability of models this year. Some little thing annoying you about some software, some feature missing just point AI at it and burn some tokens. (ok, maybe a few rounds of telling the model to do better are still needed from time to time)

vmconnect.exe (Hyper-V enhanced session) not going fullscreen after login when the window is fullscreen and you have to minimize and maximize to get fullscreen? Have your agent debug the issue and come up with an elaborate hook system that leaves the Microsoft files untouched.

Codex/Chatgpt electron app coming with annoying or missing features (Mini, Invite friends, no mcp hotreload, windows updates failing, ...)? Just point codex at codex and have it build an mcp Ouroboros to improve itself and enable dynamic js userscripts.

Deskflow clipboard sharing between devices missing file sharing, just have sessions on Mac, Linux and Windows collude to cook up a solution.

Yet writing proper documentation and reports that are human digestible still seems a bit out of reach, why would humans need to read anyway...