Have you tried ChatGPT’s new Agent Mode? It can now browse websites and click around using its own browser and mouse pointer.
We recently tested the feature to see how it performs in a real-world task that involves live websites and small decision points. The example we used was checking how many homes in Florida listed on Trulia have exactly three bathrooms. It was a clean way to see how the tool handles browser filters, small logic steps, and handoffs.
Step 1: Set the instruction
Enable Agent Mode, by typing in the command ‘/agent’. That’s ‘slash’ & ‘agent’ in one word. A prompt to select Agent Mode will appear. Go for it.
Once it’s in Agent Mode, we prompted it with:
“Please browse home listings in Florida at https://www.trulia.com/FL/ and give me a total count of homes that have 3 baths.”
The prompt was short and specific. We kept the request clear so the tool could stay focused.
Step 2: Watch how it works
ChatGPT opened a virtual browser session and began navigating the Trulia site on its own. It reached the filters without help and identified that the site did not offer a “3 bathrooms only” filter. The options were “3 or more” and “4 or more.”
It solved the gap by pulling both numbers and subtracting one from the other. That gave it the correct count of homes with exactly three bathrooms. It handled that reasoning step on its own without asking for more input.
Step 3: One small handoff
At one point, Trulia ran a verification check to confirm the session was not automated. We stepped in briefly to complete the check, then handed control back. ChatGPT resumed without issue and completed the task.
Why this is worth sharing
The task was straightforward, but it showed that Agent mode can follow a process across multiple steps, deal with missing filters, and use basic logic to reach an answer. That matters when you’re dealing with live websites where automation needs to pause, adjust, and continue.
When this is useful
You can apply this to:
- Quick data pulls from public websites
- Internal walkthroughs or SOP recordings
- Task delegation where small decisions are involved
- Scenarios where you want automation with a fallback
We recorded the session (8.5 minutes) using a screen recorder as we usually do when recording workflows.
Tools like this are going to put pressure on how remote work gets done, especially in roles that rely on repetitive tasks.
Instead of putting our head in the sand, we’re testing how to work with it, so we can move faster, stay accurate, and bring more value to the clients who rely on us.
Hope this helps!
Let’s keep it tight, keep it moving, and get real work done.
