Half this month's agent launches are "the same coding agent, but not in the terminal". What do you actually give up by leaving the terminal? Worth switching?

On question 3, from having tried the notification approach for a month: it half-works, and the half that fails is predictable.

What it fixes is real. Not knowing is worse than waiting. A notification that says "this is blocked on you" reliably saves the twenty minutes you described, and that alone justified it for me.

What it does not fix is that you now get interrupted by every run instead of checking on your own schedule. Two agents going at once and you are being paged constantly. I ended up wanting the opposite of what I asked for: not "tell me the moment it stops", but "tell me if it has been stopped for five minutes", which almost nothing offers.

So before you buy on the notification promise, work out whether you want an alarm or a summary. They are sold interchangeably and they are not the same thing.

Question 1 is the one to lead with, and there is a clean test for it.

Ask whether it drives a real terminal — an actual pseudo-terminal running the same binary you would run yourself — or whether it talks to an API and renders its own view. Both exist in this batch and they behave completely differently the moment something is unusual.

A real terminal underneath means anything that works in your shell works there: your config, your environment, your credential helper, the odd prompt that expects a keypress. A reimplementation means a curated subset, and you find the edges by hitting them at the worst moment — usually an interactive prompt the wrapper cannot forward, leaving the session wedged with no way to answer it.

The wrappers that describe themselves in terms of real terminals and session resume are telling you which kind they are. That language is a deliberate signal to people who have been bitten.

Question 2 is worth testing on day one of any trial, because it is where the thin wrappers fail and it is trivially checkable.

Start a session in the wrapper. Close the wrapper. Open your normal terminal and try to resume that session. If the agent's own resume finds it, the wrapper is a view onto real state and you have lost nothing. If the session only exists inside the wrapper, you have taken on a dependency for every session you start in it.

That second kind is not disqualifying — plenty of good tools own their state — but it converts "I could stop using this tomorrow" into "I would lose my history", and you should know which one you agreed to.

One angle missing from your list, as the person who dictates rather than types: some of these are not about supervision at all, they are about input.

The voice ones are solving "I cannot type comfortably for an hour", which is a different problem from "my agent is blocked". They just look adjacent because they wrap the same tool.

If you have no input problem, that entire branch of the category is not for you, and evaluating it against your multiplexer will make it look pointless. It is not pointless. It is aimed at somebody else.