Send an AI shopping: agent mode, explained honestly
There is a moment when AI stops being a chatbot: the first time you watch one open its own browser, visit a store, and do something for you while you watch. The viral version of that moment is beautifully mundane. It hunts discount codes.
The trick runs on ChatGPT's agent mode, which OpenAI ships for paid plans: the AI gets its own browser, narrates what it is doing on screen, and asks permission before anything consequential. Here is the trick, and then the part the viral reels skip.
Screenshot the thing you want to buy
Any product page, any store. The screenshot tells the agent what the item is and where it lives.
Give the agent its mission
Find me a working discount code for this item. Go to the store's site, add the item to the cart, and test codes one by one until one works. Show me proof of the discount. Use agent mode. Do not enter any payment or personal information.
That last sentence is ours, not the viral version's. Keep it.
Watch it work, then take over
The agent opens its browser, finds the item, tries codes in the cart, and reports back with the winner applied. Then you step in for the only part that should ever be human: paying.
One: the agent never gets payment details, saved cards, or autonomy to purchase. Two: you review everything it did before acting on it. Three: expect failures; some sites block automated browsers, and some days there is no code. The agent saves time. It does not replace judgment.
One distinction worth holding on to, because a second kind of browsing AI now exists and it is not this one. Agent mode opens its own fresh browser, signed in to nothing — which is why rule one above is enforceable at all. Claude in Chrome does the opposite: it drives your browser, with your existing logins, on the page you are already on. Safer in one direction, far more capable in the other. We broke that one down in what Claude in Chrome actually does, and what to keep it away from.
Questions people actually ask
Is this free?
No. Agent mode requires a paid ChatGPT plan (Plus, Pro, or Team). The technique itself is one prompt, so the only cost is the subscription you may already have.
Does it always find a working code?
No, and be suspicious of anyone who says otherwise. Sometimes there is no valid code, and some shops block automated browsing entirely. It wins often enough to be worth one minute of waiting.
Can it buy things without me?
It is designed not to: OpenAI's agent asks permission before consequential actions, and you can interrupt or take over at any point. Our house rule goes further: never store payment autonomy with any agent. The agent finds; you pay.
Is this the same as the AI agents businesses use?
Same species, consumer edition. The pattern (an AI that acts instead of answers) is exactly what we build for companies: agents that chase invoices, fill spreadsheets, and follow up leads while you do something better.
Want this built for you?
We write these memos because we build this stuff every day. If you want it working in your business instead of sitting on your reading list, that is literally our job.