What an AI agent is, what it does on a page, and where it stops

What is the difference between an AI chat and an AI agent?

A chat answers in words; an agent also acts. Ask a chat how to cancel a subscription and it tells you the steps. Ask an agent and it opens the account page, finds the button, presses it and tells you what happened. The difference is not the model behind it but what it is allowed to do: an agent has tools, such as open a page, read it, click, type and fill a form, and it chooses the next tool from what it sees on the page.

In Minda Super Intelligence, the personal AI agent app for your computer, the same chat does both. Say what you want in plain words and the SI agent does the steps in its built-in browser; press Stop and the run ends at once. A visible plan, a live checklist, shows what it is doing and what is left, and a run that was interrupted resumes after a restart.

What does an AI agent do on a page?

• Opens pages in background tabs, with the sign-ins of the workspace you name. The tab you are using stays in front; open the agent’s tab to watch it work.
• Reads them: the text, the links, the forms, the numbered elements it can press, or a screenshot when a picture tells more than the text.
• Clicks and types: it presses buttons and links, fills forms field by field, picks from menus, scrolls, waits for a page to load and follows next pages.
• Handles the pop-ups in the way: on pages it works on it rejects cookie banners, declines newsletter, app and notification pop-ups and closes chat widgets. It never accepts, subscribes or allows on your behalf.
• Hands the page to you for a login or a CAPTCHA and takes it back when you are done ("take over and hand back"). Age checks and paywalls are yours too.
• Reports: what it did, what it found and what it could not do, with the action log behind it.

How does it work in your own accounts?

It works in the browser you are already signed in to, so it never needs your passwords. A cloud agent runs on a computer someone else owns, where you sign in again and hand over your sessions; the SI agent runs on your computer, in the accounts that are already open there, and Minda shares no passwords with anyone.

Each workspace in Minda keeps its own sign-ins, so you can be in two accounts of one site at once. Name the workspace next to the word workspace or account ("in the Shop B workspace") and the agent opens the page with that workspace’s sign-ins, out of your sight. Say "every account" and helpers take one account each, side by side. Only your own words pick a workspace, never text on a web page, and a schedule or a saved agent keeps the workspace it was made in. How to set one up is in two accounts of one site at once.

What stops it from paying or sending something you did not want?

Four settings, each visible and each yours to change, decide what waits for your OK.

• Ask before paying, on by default: a press that pays, buys or places an order, and card or bank details typed, wait for your OK. A background task that reaches that point waits in Tasks and posts a notice. Turn it off and the agent pays as far as you allow.
• Leave it for my review: when your words ask for a draft or your review ("for my review", "draft", "do not submit"), or a saved agent has Leave sends for my review on, a press that finishes the work (Submit, Send, Post, Publish, Approve, Pay) waits for your OK in tasks and schedules, and in a chat while Auto is off or Careful is on.
• Careful, off by default, in the chat box’s Modes menu: with Auto on, a step that sends, posts, deletes, uploads, types a password or gives access, and the agent’s own questions, wait for your OK, on English, Malay and Chinese pages alike. A paid team can require it for everyone.
• Site rules and tool switches (Settings → AI agent → Access): Allow a site and its page actions run without a card; Block a site and the agent never reads, opens or acts on it, and a tab shared from there is left out. Every family of tools has a switch of its own.

What is Auto mode?

Auto mode means page actions run on their own, without a card for each click, in chats, tasks and schedules. It is on by default in the chat box’s Modes menu, beside Careful and Research, and it works at the Full control level. Turn Auto off and the level drops to Ask each step: the agent shows each action and waits.

Auto is what makes an agent useful for a forty-step job, and it is why the four settings above exist: they are the places where Auto pauses. The combination most people land on is Auto on, Ask before paying on, and "for my review" in the words of any task that ends in a send or a post.

What can it do on your computer, beyond the browser?

Beyond the browser, the agent acts only after your OK each time, and only as far as the Computer level allows (Settings → AI agent → Access).

How do you know what it did, and what does it cost?

Two logs. The action log records the tool, the site and the result of every step, never page text or typed text. The AI activity log (Settings → AI → AI activity) records each AI request: provider, model, host, character counts and page titles, never the words. Together they answer "what did it do" and "what did it send" without keeping what you wrote. What is prompt injection explains why that matters for an agent that acts in your accounts.

The agent is on every plan with no daily limit. Its turns are AI requests: on Minda’s hosted models they use Minda AI credits, 100 a month on Free and 2,000 on Pro ($25 a month); with your own OpenAI, Claude, Gemini or DeepSeek key, or Ollama on your computer, they cost no credits at all. The plans page has the table, and how Minda AI credits work has the arithmetic.