Why it behaves how it does

Voice mode is built to sound like a person talking to you, not a document being read aloud. A few things it does on purpose are worth knowing, so they don’t catch you out.

On GPT-Live, there are two of them

This is the thing to understand. On a GPT-Live call you’re talking to a voice model built for conversation: it listens while it speaks, it can be interrupted, and it’s quick. What it can’t do is open your notes, run a task, or use any of your agent’s tools.

So it works with your agent. When you ask for something that needs real work, it hands the request to your agent in the same conversation, keeps talking to you while your agent does it, and tells you what came back in its own words. You never ask it to do this; it decides, the same way you’d expect a good assistant to know when to go and check.

A few things follow from that:

• Small talk is instant, real work takes as long as it takes. “How’s it going?” comes straight back. “What’s on my plate this week?” means your agent goes and looks, and the voice tells you when it has it.
• It might say it’s still on it. If your agent is taking a while, the voice checks in rather than leaving you in silence. That’s normal.
• What you hear is the voice’s version of what your agent found. The transcript keeps both: your agent’s work, marked as handed off, and what the voice actually said. Nothing your agent said is hidden; it’s just not read at you twice.
• The call starts knowing the conversation so far, the date, the time, and what’s on your screen. It doesn’t start knowing everything in your vault. Anything it doesn’t know, it asks your agent for.

It says the short version out loud

If you ask something whose answer is a long list, a draft, or a page of detail, it won’t read all of that aloud. Instead it tells you the short version out loud and puts the detail on screen for you to read. Lists, numbers, tables and drafts all land there.

So if it gives you a brief spoken answer and you were expecting more, look at the screen. The full thing is usually there.

The whole call is in the thread

Every spoken turn, yours and its, is written into the conversation as it happens, marked as voice. Pick the same tab up by typing later and your agent knows what was said on the call, so you can carry on from where the conversation got to rather than starting again.

Why it’s built this way

The whole point is that it talks the way a person would. A person answering you out loud gives you the gist and shows you the document, rather than reading every word of it at you. And a good assistant on the phone doesn’t go silent while they look something up; they tell you they’re looking. Voice mode does the same, so listening to it feels natural rather than like being read a report.