Voice, browsing and agent features

Voice is for thinking out loud, browsing is for fresh facts with links, and agents act for you. The more a feature can change, the closer you watch it.

Voice: hands-free thinking

Voice mode lets you talk to ChatGPT and hear it talk back. It is good for:

Rehearsing a difficult conversation on the drive in.
Brainstorming while you stock shelves or walk between jobs.
Dictating a messy first draft you will clean up later.

It is less good for anything with numbers or names you need to get exactly right, because you cannot see them. Switch to text to check. And remember where you are: a voice chat about a customer in a café is a privacy problem no setting can fix.

Browsing: fresh facts with links

ChatGPT’s built-in knowledge is frozen at a point in the past. When it browses, it searches the live web and answers with links to what it read.

Use browsing when the answer depends on something recent: prices, regulations, product details, opening hours, news. Then do two things every time:

Open at least one link. Does the page actually say what ChatGPT said?
Check the date. A 2022 page about fees is history, not fact.

If a number appears with no link, and asking “what is the source for that figure?” does not produce one, treat it as a guess.

Voice, browsing and agents each come with their own habit:

Three cards: voice for thinking out loud, browsing for fresh facts you click to check, and agent mode for tasks sorted by zone — Three features, three habits. The agent habit is the three-zone rule below.

Agents: from answering to doing

An agent does not just answer. It takes actions: browsing sites, filling forms, comparing options, completing multi-step jobs, sometimes in a browser it controls itself. It can save hours. It can also do the wrong thing quickly and politely.

Sort every agent task by what it can change.

Zone	What the agent may do	Your role	Examples
Green: read only	Look, gather, summarise	Let it run, check the result	Compare five suppliers’ prices into a table; find the deadlines in a set of PDFs
Amber: reversible	Draft or create things you can undo	Watch, approve steps, review before save	Draft replies; fill a form but stop before submit; build a spreadsheet
Red: irreversible or external	Send, pay, delete, post, sign, agree	Do the final step yourself	Email a customer; place an order; delete records; post publicly

An agent should not be in the red zone alone. Not because it always fails, but because when it does you cannot take it back.

The three zones at a glance:

Three zones for agent tasks: green read-only runs alone, amber reversible changes are watched, red irreversible actions stay with a person — Sort by what the agent can change, not by how clever the task is.

How agents go wrong

Wrong target: perfect form, wrong customer record.
Confident fill: a field it cannot find gets a plausible value instead of a question.
Scope creep: asked to tidy one folder, it tidies three.
Instructions on the page: websites can contain hidden text aimed at AI agents. Good agents resist this; none are perfect. Keep agents away from payments and passwords.

Briefing an agent

Say what done looks like: “Draft replies to the three supplier emails. Do not send.”
Set a boundary: “Only use the Invoices 2026 folder.”
Stop before the irreversible step: “Fill the order form and stop. I will click submit.”
Watch the first run of any new task. Earn trust before you grant it.

How it looks today · ChatGPTChecked Nov 5, 2026

Voice: tap the microphone or waveform icon beside the message box. Tap it again or say “stop” to end. The transcript appears in the chat afterwards.

Browsing: ChatGPT often searches automatically when the question needs it; you can also choose a search option from the tools or + menu. Sources appear as link chips or footnotes; tap one to open it.

Agent mode: on paid and workspace plans, look for an agent or “do this for me” option in the tools menu. It narrates its steps and asks for confirmation before certain actions; it may also pause and hand control back to you to sign in or confirm. Keep those confirmations on. Admins can restrict agent features on workspace plans.

Menus and button names change often. If your screen looks different, the skill is the same: look for the equivalent option.

Two short clips from OpenAI. First, voice in everyday use:

Video: Get Stuff Done with ChatGPT Voice · OpenAI · 2 minNotice that nothing with exact numbers or names is settled by voice alone; those still need a text check.Watch on YouTube (opens in a new tab)

Then the agent, acting rather than answering. As you watch, sort each step into a zone:

Try it now

Write down three tasks you would like ChatGPT to do for you this month. Put each in a zone: green, amber or red. For any red task, rewrite it so the agent stops at the final step and hands over to you.

Remember: voice for thinking, browsing for fresh facts you then click to check, agents for tasks sorted by zone. Read-only runs alone, reversible gets watched, irreversible stays with a person.

Checking your progress…