We’re doing a home renovation, and that means we need to clear out a lot of stuff. My wife’s Vinted webshop had 400+ clothing items that needed to go.

To make things move faster, we wanted to decrease prices on all these items. Doing this manually would be a lot of repetitive work, so with all the recent buzz around agentic browsing, I wanted to see if these tools could help.

Testing Perplexity Comet

I started with Perplexity Comet:

Visit every single one of the urls below one by one. Each url contains an edit form for an article in a webshop. Decrease the price by 50%, but do not go below 9,95 EUR and click the save button. Log the URL every time you've handled one for future reference. After each save, continue with the next url until you have handled all urls below. This will take a long time, but it is essential you do all urls. Only keep track of the urls you have handled, do not log any other information.

// list of urls

It worked for small batches, but it really did not want to run through the full list. It kept complaining the task was too time consuming. Even after multiple attempts to convince it to continue until done, it still stopped early.

Batches of about 10 URLs worked reasonably well.

Here’s Comet in action, processing a batch of URLs:

Comet: While it successfully handled individual tasks, the inability to process large batches made it impractical for our 400+ item scenario. The constant need to restart and manually manage batches defeated the purpose of automation.

Testing OpenAI Atlas

Then I tried OpenAI Atlas. Initially, I had quite some issues getting it to work. A plain prompt did not trigger Agentic mode. Once I found the right setting, it was relatively fast, but it still refused to do more than 3 items in one run, even with extra convincing.

I used the same prompt as above.

The following video shows Atlas processing a few items before stopping:

Atlas: Despite its speed advantages, the hard limit of 3 items per run made it equally unsuitable for bulk operations. The agentic mode required manual activation and still needed constant supervision.

Testing Claude Code

For Claude Code, I switched to a more explicit pseudo-code style prompt:

# Agent Instructions

You are an agent managing a webshop.

## Your Task

1. Read the urls.txt file
2. Read the progress log at `progress.txt`
3. Read the done log at `done.txt`
4. Remove the first line from urls.txt and put it in progress.txt
5. Navigate to the URL and:
    - Decrease the price (data-testid="price-input--input") by 50%. Make sure the price does not drop below 9.95.
    - Click the "Save" button (data-testid="upload-form-save-button") at the bottom of the page and wait for the new page to load.
6. Move the line from progress.txt to done.txt with the old price and the new price, separated by a comma. For example: `https://example.com/product/123,19.99,9.99`
7. Repeat the process until urls.txt is empty.

This approach could have handled the full list, but it was too slow: about 2.5 minutes per item. It also burned through tokens very fast. I stopped it after around 3 hours, after hitting both my session limit and 42 USD in free credits.

Watching Claude Code working through the task, slowly:

Claude Code: While theoretically capable of completing the full batch, the economics didn’t work out. At 2.5 minutes per item and $42 spent on just a fraction of the list, completing all 400+ items would have taken over 16 hours and cost hundreds of dollars.

Pivot 1: Puppeteer script

Given that we were already at pseudo-code level, I asked Claude Code (Opus 4.6) to generate a Puppeteer program for the full batch.

In theory this should have worked, but the webshop platform immediately flagged it as a bot, even with spoofed headers.

Pivot 2: Chrome plugin

Finally, I asked Claude Code (Opus 4.6) to build a Chrome plugin instead:

Create a Google Chrome Plugin that:

allows the user to enter a list of urls (separated by newlines)
visit each url one by one:
a. find the textfield with queryselector [data-testid="price-input--input"]
b. read the existing value in that field as a float
c. decrease that value with 50%, but keep a minimum value of 9.95
d. click on the button with queryselector [data-testid="upload-form-save-button"]
e. wait for the next page to load + an additional second
f. continue with the next url, util all urls were visited

On the first shot, it built a plugin with setup instructions. I needed a couple more refinement prompts to get it fully working:

1. The initial code assigned a new .value to the input field. But because the webshop UI is React based, it needed a click + type + blur flow to pick up the new value.
2. After the edit in step 1, clicking in the price field made the popup disappear. After prompting the popup often disappears after doing one price update. Maybe it has to do with the click on the page...? Find a way to keep the plugin UI visible & running., it figured out a technique with a service worker.

The plugin worked like a charm, and it processed the full list in about 30 minutes, without any supervision or manual restarts. The best part: it did not trigger any bot detection, so I could run it on my main browser without any issues.

Conclusion

For this scenario, agentic browsing has clear limits: long repetitive flows still tend to stop early, slow down too much, or become expensive.

The most efficient path was using AI to generate a small purpose-built tool. Instead of asking an agent to do 400+ manual browser actions, asking it to build the right automation turned out to be the better move.