There is a massive gap in reliability. In short: CLI/API agents are "Junior Engineers," while GUI agents are currently "Clumsy Interns."

If you are interested in technical efficiency (Linux, Vim, etc.), you will likely find the GUI agents frustratingly slow and error-prone compared to the CLI workflows you are used to.

Here is the data-backed comparison:

GUI Agent (Computer Use) - OSWorld (General desktop tasks) | ~22% - Low: Fails 4 out of 5 times on complex tasks.

CLI / API Agent (Tool Use) - SWE-bench (Software Engineering) ~50 - 62% - High: Can reliably handle complex logic half the time.

GUI Reality: On the OSWorld benchmark (which tests things like "Open Excel and plot this data"), Claude 3.5 Sonnet initially scored 14.9%, improving to 22% with better prompting.

Compare it to Humans score ~75%.