Yesterday, OpenAI announced ChatGPT Atlas on a livestream. They promise a new AI-native browser that allows you to finally go beyond AI summaries and advise but have their agent actually complete tasks for you.

Sounds too good to be true, so I immediately jumped on it and spent the last 24 hours testing it. In today’s post, I share:

Let’s jump in!

Lots has been said in the launch stream. Here’s the TL;DR of what’s new and promised:

Go to chatgpt.com/atlas. You must be using a Mac, click download, and then follow these simple steps:

Once downloaded, you’ll be guided through a simple onboarding flow, starting with your ChatGPT login credentials, supported by SSO.

It then allows you to import your browsing history, passwords, and bookmarks from previously used browsers.

Browser Memories allows ChatGPT to search your activities and further optimize your experience through personalization - claiming that all data remains private. Hopefully..

I skipped the data import and browser memories, so started almost on a white sheet of paper except for my previous ChatGPT history.

Opening a new tab, you always start with a console that takes both regular websites but also prompts to execute tasks. So there’s a unified starting point taking you in various directions.

Landing on a website, you can easily open the sidebar with “Ask ChatGPT” and then interact with the website. Some examples in the screenshots below + the most valuable prompts in section “4. Top Prompts” of this post.

Browsing the web for few hours, I already hit the first pages blocking the ChatGPT sidebar from scraping and processing their content, e.g. NYT👇

All in all, I really enjoy the option to directly interact with pages and process their content for summaries, comparisons, etc without switching tabs. It’s more convenient then jumping back and forth across tabs, saving few seconds for every step - which can quickly compound to minutes and hours.

Let’s get to the fun stuff: Agent Mode - promising to “actually get work done for you”. While the concept is not new and initially got released as “Operator” by OpenAI in January this year, this capability now takes center stage in a major product release, marking a big step toward bringing such autonomous systems directly into the hands of end users.

Note that right now it’s only available in “preview for Plus, Pro, and Business users”, so let’s get a first impression by automating a common task.

Login to LinkedIn, go to Sales Navigator, and ask the agent to create a search for people who work at “Stealth Startup” and are based in Munich. It’s a simple one-step search automation. Here’s the result.

Subsequently, I increased the complexity with a multi-step automation by asking the agent to 1. screen profiles, 2. go to profiles of every person who is less than 6 months in stealth, 3. learn more about what they do, and subsequently 4. create a list with Name, Title, Company, and a brief description about what the person builds.

While the automation is relatively slow compared to other browser based automation tools like PhantomBuster, Bardeen, BrowserUse, Zapier etc., it’s pretty nice to follow the flow and have the option and jump in where needed. You can see below how the curser has a small visual tag describing what it does at every step, making it easy to follow.

The accuracy across the multi-step automation is surprisingly high for this more complex flow (and also others that I tested). It does exactly what it’s supposed to do, no more and no less. Below is the result of the multi-step automation visible in the sidebar.

Additionally, I’ve tested more complex tasks that are a mix of repetitive automations and reasoning. Browsing LinkedIn, commenting on posts, processing my inbox and drafting email replies. Yet, there are some limits when it comes to write functionality that requires reasoning. The limitations here are the same as for ChatGPT standalone. Sometimes it just takes the wrong turn, forgets relevant context, or requires human intervention.

Share this article with others who might benefit.

Here are some further prompts for startup sourcing, screening & due diligence, and deal winning & closing. Find our full prompt library here.

After testing Atlas for 24 hours, here’s the summary of my first impression.

The good: The experience of browsing feels smoother: instead of bouncing between search engine → results → site → back, you type a query or highlight content and let the assistant handle context. One key shift: I found myself not opening Google as first reflex, but instead typing into the built-in ChatGPT prompt. Atlas essentially positions ChatGPT as the default surface for queries and navigation.

The sidebar assistant is very capable and super useful for investors as we browse the web and then drill down on company pages with further analyses and research. The sidebar allows you to do it all without context switching.

More importantly, the Agent Mode is a game changer. I’ve previously used various workflow automation tools like Zapier, n8n, Make, Bardeen, BrowserAutomate, and more, but now a lot of their functionality can be done natively in Atlas - without the need for additional tools. It’s still relatively slow and functionally a bit limited, but accuracy is high - driving trust, the most critical when it comes to write-automations.

Another feature that’s available but I didn’t yet test is the browser memory search. Investors have so many things in their mind and startups on their radar that too often we struggle to recall where we found what. If you’ve ever struggled to find an article you opened two weeks ago, now you can ask: “what was that machine-learning due diligence piece I read last month?” and the assistant surfaces it. That utility alone feels like a significant productivity win for heavy users but I’ve been afraid - at this point - to open my full history and share all my data. Potentially in the future..

The caveats: The most obvious, access is limited to macOS (for now) and pay-walled for Plus, Pro, and Business users. On the usage itself, some sites actively block side-bar interaction or scraping for summarisation. That’s a bummer and reduces Atlas to regular browser functionalities.

What’s also interesting is that I cannot login with multiple Google accounts, only the one that’s connected to my ChatGPT subscription. Obviously, that limits SSO and breaks flows for various logins.

Most importantly, privacy and data risk still loom. Even though memory is optional and data usage is opt-out by default, giving a browser deep access to your tabs, history and tasks creates a new attack surface. Analysts are already warning about “prompt-injection attacks, data exfiltration, or the browser being turned against the user” in the context of AI-powered browsers. Granting access to such critical data requires a lot of trust and given that Atlas is only available for 24 hours, it might take more time to see how robust it is against attacks, and how personal data is actually handled.

Conclusion: It’s early but I like what I see. The side bar is more of a convenience factor that removes the need for tab/context switching. The agent mode is the true differentiation. While I’m not yet at “set it and forget it”, it’s already strong for simple, repetitive tasks that humans can spot-check afterwards. As I gain more trust (hopefully) over time, I might grant access to more data and allow it to perform more complex tasks. I’ll keep you posted.

Stay driven,
Andre