I already wrote about it in my piece on AI agents that browse the web: web agents are more and more popular. They can carry out actions on the web in a window, fully autonomously.
An example:
In terms of interface, there is a conversational chat on one side and the executions on the other.
There are quite a few limits to this approach. Because the executions run in the cloud, the agent often lacks context. It is up to us to share all the information it will have available for its actions.
This is where web agents built into a browser make sense. And that was Perplexity's bet with the July 2025 launch of Comet: rethinking the browsing experience in the age of AI.
We are used to using our browser to click on blue links and end up with dozens of open tabs. We get lost and often want a summary of some content, to extract information, to compare information across pages, and so on.
Our browser looks like dozens of open windows.
Comet's arrival lets you pull tabs into a conversational chat, and therefore share context to get an answer.
You can get a summary of a page in 1 click, or discuss the content of a YouTube video to find out what it covers.
Where I notice a big difference is in the integration with our tools. When I install Comet, I can import all the information from my Chrome profile (cookies, passwords, etc.). This means being already logged in to every tool: Gmail, Notion, Amazon, and so on.
No more need to share extra information (including passwords) with the models: it opens a window and performs actions while already logged in to my tools.
For example, it lets you add a conversational layer to every tool I use. I can ask which newsletters I receive by email, then unsubscribe from some of them. Or move an invitation and send an email (while keeping a human validation for sending the email).
I made a YouTube video where I share 5 concrete examples of using Comet and what really changes compared to classic web agents.
This access to all my tools has a downside. In August 2025, Brave showed that a malicious web page could hijack Comet's agent through prompt injection. That is the whole challenge of AI agent security in the enterprise.
Alternatives to Comet as a browser
Other alternatives exist, such as Dia Browser, which I found more limited in July 2025, because you could not carry out actions; it was more about adding AI features that improve the experience.
OpenAI releases ChatGPT Agent to keep up
In January 2025, OpenAI had released Operator, a web agent able to carry out autonomous actions, but only available on the $200/month Pro plan.
In July 2025, they went further with ChatGPT Agent, which combines deep research and Operator.
The solution is a combination of several tools OpenAI already has, but it is frankly not revolutionary compared to what is already on the market for web agents or browsers like Comet.
The big plus is OpenAI expanding its features into an ecosystem of increasingly powerful tools. But nothing new under the hood based on a quick analysis.
I get the impression the tool also has a slightly larger context window, letting it carry out more complex or longer tasks, such as the ability to create files, store them on cloud services (Google Drive for example), create slides, or fill in a Google Sheet after having done a web search beforehand.
An example here on slide generation (still in beta):
October 2026 update: Perplexity made Comet free for everyone in October 2025. The same month, OpenAI launched its own browser, ChatGPT Atlas, then shut it down in August 2026 and folded browsing into its ChatGPT desktop app. For high-volume processes, my view on web agents hasn't changed: an automation plugged into APIs is still faster and cheaper than a browser that clicks.
