Why I believe in autonomous agents: a ChatGPT agent case

OpenAI is trying somehow to smooth over the questionable launch of GPT-5. They’ve already brought back access to GPT-4o and let you choose GPT-5’s mode yourself (auto, manual, thinking). I hope they manage to fix the rest of the shortcomings soon. But that’s not what I want to talk about today. I’ll share an interesting case that clearly shows why I believe in autonomous agents so strongly.

Recently I went to ChatGPT with a request to visualise the statistics of IronMan 70.3 participants by finish time. GPT-4o searched the internet and kindly gave me statistics for the top 24 participants. I insisted on extending it to all participants. But GPT-4o couldn’t find complete data on the internet. Instead, it found a site with the statistics I needed, shown as a frequency chart. And it suggested that I click on every bar, write the values into a table it had prepared, and then it would format everything nicely for me. :)

I wasn’t ready for that exercise. I switched on agent mode and asked it to finish what it had started. And that’s where things got interesting. The agent opened the site in a browser, easily dealt with the pop-up ads and with the filters for showing the statistics I needed. It got the chart and started trying to click on the bars. But the bars all have different heights, they’re thin, and some are very low. Even a human would struggle to click on some of them.

It clicked long and patiently until it reached one very tricky bar. It made five attempts and failed to click on it. Then it gave up on that approach and decided to look for the bar data in the page’s HTML code. It worked out what to hook into and found where in the code the data was bound. But the data was loaded dynamically, so it decided to go back to clicking on the chart bars. This time it managed to go through all the bars, using the technique of clicking at the base of each bar.

Then it decided it would be great to verify this data, since there was a risk it had mis-clicked somewhere. So it went back to the page code and found the JS code that loaded the data dynamically. Then it wrote similar Python code and used all the cookies from the browser. It got the complete data, checked it against what it had collected from the chart, and confidently produced the report I needed.

What particularly surprised and pleased me in this case:

• the persistence with which the agent worked on the task;
• the variety of techniques and tools it used;
• a great chain of reasoning, better than most people’s;
• self-control and double-checking of risky data.

What interesting tasks have you solved with ChatGPT agent?