Author(s): Felix Pappe

Originally published on Towards AI.

What if your browser didn’t just open pages, but did the work for you?

For the last month, I made Perplexity’s Comet my default browser and treated it like my search assistant.
I asked it to summarise long YouTube videos, plan a one-day walking tour through London, hunt for cheap flights with specific time windows, and draft real emails.

The results were eye-opening. I discovered where an agentic browser genuinely saves time, where its screenshot-driven workflow still struggles, and why so many AI companies suddenly care about owning the browser itself (hint: data access and lock-in matter more than you think).

In this post, I share my experience with the Comet Browser and the future of the agent-based internet, including honest case studies and concrete pitfalls to avoid.

If you’ve wondered whether an AI-powered browser can actually replace your current setup, let’s find out.

My favourite Feature of the Comet Browser

Comet offers many additional features compared to a traditional non-agentic browser. My favourite and most useful feature over the past month is something I want to share in this section.

Prompt Suggestions

Your Comet assistant already provides suggestions on what it can do for you on the current website.

For example, if you have opened a PDF document in your browser, it suggests providing a summary, without entering any addtional prompt.

Image Snipping and Audio Dictation

By clicking on the dotted rectangle in the prompt bar, as shown in the picture above, you can take screenshots of specific parts of your screen, which are immediately added to the agent's context window.

This tremendously improved my workflow and made the process of providing visual information to the model much easier.

Another feature I often use is the Dictation tool, represented by the microphone next to the Snipping Tool.

I used this tool when I was too lazy to write out my prompt completely and only wanted to get some information quickly by speaking out loud my thoughts.

Saving frequently used Prompts

A feature I really like is saving frequently used prompts, which has proven to be effective. Prompts can be saved with just one click on the square with the diagonal line at the end of each generated answer.

Then, a new window opens, allowing us to modify the prompt further and select the model and its search type for future inference.

Now AI can run this prompt every morning just by entering /ai-news-summarizer-today, instead of typing out the prompt or copying and pasting from a different source.

Case Studies

In these case studies, I provide more information about my experience in the last month about Comet’s abilities to solve specific tasks on its own without any human interaction.

I propose a selected task, which goes beyond. “Can you explain this or that?” or “Provide more details on this topic”, because we all know how well Perplexity can look up information on the internet.

Summarizing Videos

YouTube videos are often quite long, and I don’t always have time to watch all the videos I'm interested in.

So instead of watching them all. I started generating a summary of the most intriguing insights before watching it.

After generating the video summary and scanning through it, I often decided not to watch the video.
Either because the summary was enough or because it is not worth my time to watch the video, based on the brief assessment of the summary.

In the following image, you can see an example of how such a summary can look.

Planning City Trips

I must admit that summarising YouTube videos is a straightforward task that can also be accomplished with other tools.

Let’s continue with a more challenging task, such as creating a city walking tour through London.

I used the following prompt to create the tour.

I have one free day in London. Please provide me with a walking city tour to see the main attractions.

The answer was generated in a few seconds.

Afterwards, I asked to visualise the route in Google Maps, using the following simple prompt.

Show me this route on Google Maps

The generation of the final visualisation took not less than 10 minutes.

The reason for this long processing time is rooted in the working principle of the AI agent.

The agent utilises the graphical user interface in the same manner as humans do.

It takes a screenshot, analyses the image, generates an instruction, executes the instruction and repeats this process until the previously defined goal is achieved.

Depending on the task's complexity, this may require multiple iterations, resulting in extended processing times.

This process is repeated iteratively until the final outcome is achieved. In this case, the visualisation of the walking tour through London.

From my point of view, the result looks ok. However, I would think a local who lives in London would come up with a better route.

Searching for Flights

The search for flights in the past was always the most annoying part of organising trips for me.

For this reason, I hoped to outsource this task to my AI-agentic browser assistant in the future.

So, I instructed perplexity with the following prompt, including the required information.

Please look for a cheap flight from Berlin to London on February 19, 2026, and a return flight on February 22, 2026. I would like both flights to depart between 10:00 am and 8:00 pm.

The agent initially used Google Flights to solve this task and later switched to Skyscanner.

As already mentioned, it takes a considerable amount of time for the agent to navigate the user interface's design, which is not optimal for machines.

Entering even a single piece of information into an HTML form can be a time-consuming process. Thus, filling out an entire form can take several minutes, depending on its complexity.

Ultimately, the agent produced the following output, showing a list of cheap flights from Berlin to London.

While this output appeared fine at first glance, upon investigating the links, it became clear that the agent only considered London City Airport and ignored all other airports in London, which offer cheaper prices.

The entire process took seven minutes, without the time which I needed to fact-check the information.

When I manually entered the information into Google Flights, it took me less than a minute.

Answering Mail

One task on which most of us spend a lot of time is writing emails.

For this reason, I also tried to write and respond to a few emails with Comet.

In the following example, I used this prompt for drafting an email for postponing a meeting.

Write an email to Max Mustermann (max.musterman@gmail.com), stating that we have to postpone our meeting this Friday to next Friday.

Most of the time, I was satisfied with the outcome generated by the assistant. The generated email output for the query above is as follows.

Hi Max,
I wanted to let you know that we have to postpone our meeting this Friday to next Friday. Sorry for any inconvenience caused. Please let me know if the new date works for you!
Best, Felix

However, I learned that I must describe the task more thoughtfully when instructing an AI assistant that takes action on a real-world system.

It happened to me too many times that the agent immediately sends the email without providing me the chance for a proofread.

So don’t forget to add “Don’t send the e-mail, just write it” to prevent this misunderstanding.

Personal Thoughts after one Month

After switching browsers and trying an agent-based one for the first time, a lot fell into place.
In this section, I’ll share those takeaways, especially how it made recent news coverage of Chrome clearer: the debate about a possible split from Google and Perplexity’s reported interest in acquiring the Chrome browser.

Merging both Worlds

Using Comet Browser from Perplexity as my default browser for one month made clear to me why so many companies are currently interested in Chrome Browser or building their own.

The browser has evolved into the “natural” user interface we have grown accustomed to over the last few decades, for looking up information and using the internet.

With Comet as my default browser for the last month, I used AI-generated answers more frequently than ever before.

This made me aware of how much the proper integration of AI agents into the browser interface will drive the future adoption of AI.

Merging the old internet world with the new agentic AI world enables agents and their usage patterns to be integrated more smoothly into existing workflows and broader society, including all their advantages and disadvantages.

Data Access

Moreover, browsers enable AI companies to collect new data that they were previously unable to access.

For example, consider The Verge. I can ask the Comet assistant about the most recent AI news based on The Verge's website. As you can see in the screenshot, it worked perfectly.

However, if I do the same through the perplexity API, The Verge does not allow me to get the information.

This highlights that an AI-agentic browser allows AI companies to access content that they previously could not, because I open the content on my machine in my browser, not through a bot.

Browser Lock-in effect

Another big lesson from this experiment is that we’re far more dependent on our browsers than we realise, and switching is hard.

I’ve used Chrome for years and customised it to my workflow. Without realising it, it had become woven into my daily routines, which made moving away surprisingly inconvenient.

I hadn’t realised how locked in I was to Chrome; switching from it to Comet almost felt like changing to a different operating system.

Final Results

Now the question arises: will I continue using Comet as my default browser?

The answer is simple: No, not yet.

While Comet has some great features I don’t want to lose, such as the smooth integration of video summaries and PDFs. However, these aren’t exclusive to the browser. Other tools can also provide them.

Its main feature, the AI assistant, can perform specific tasks independently. Although the agent generally performs well and can solve most tasks, it’s limited by its iterative, screenshot-based workflow.
In almost all scenarios, the agent was much slower than doing the work manually.

Now it can be argued that you can run multiple agents in different tabs, but my laptop really struggled with that because it requires more resources than normal browsing, even though the model isn’t running locally.

In conclusion, it’s impressive how well the AI agent can navigate the web using screenshots of the user interface.
However, it’s too slow to improve my workflow meaningfully and is currently more of a novelty than a productivity enhancer.

That said, this doesn’t mean we won’t see further improvements. I’m optimistic that we’ll see major advances in how we interact with the internet through AI over the next few years, one of the most promising paths to the broad adoption of AI models.

What are your thoughts about the agentic browser? Have you already tried Comet Browser or other agentic tools?

I would love to hear your thoughts in a comment below.

Best Regards
Felix

Follow & Connect

https://felixpappe.substack.com/ 📧
felix-pappe.medium.com/subscribe 🔔
www.linkedin.com/in/felix-pappe 🔗
https://felixpappe.de🌐

P.S.: If you’ve read this far, I hope you’ve gained some value from this article. If you would like to learn more about AI, I would greatly appreciate your support of this project by subscribing or providing feedback to generate more valuable posts about AI for people interested in this recent technological advancement.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI