“Agentic” is the buzzword to distinguish “LLM will tell you how to do it” versus “LLM will just execute the commands it thinks are right”.

Particularly if a process is GUI driven, Agentic is seen as a more theoretically useful approach since a LLM ‘how-to’ would still be tedious to walk through yourself.

Given how LLM usually mis-predicts and doesn’t do what I want, I’m no where near the point where I’d trust “Agentic” approaches. Hypothetically if it could be constrained to a domain where it can’t do anything that can’t trivially be undone, maybe, but given for example a recent VS Code issue where it turned out the “jail” placed around Agentic operations turned out to be ineffective, I’m not thinking too much of such claimed mitigations.