Run web searches and browser tasks with Operator
I am the weirdest AI fanboy you'll ever meet. Task: I am the weirdest AI fanboy you'll ever meet. I've used every single major large language model you can think of. I have completely replaced VSCode with Cursor for my IDE. And, I've had more subscriptions to AI tools than you even knew existed. This includes a $200/month ChatGPT Pro subscription. And yet, despite my love for artificial intelligence and large language models, I am the biggest skeptic when it comes to AI agents. Pic: "An AI Agent" — generated by X's DALL-E So today, when OpenAI announced Operator, exclusively available to ChatGPT Pro Subscribers, I knew I had to be the first to use it. Would OpenAI prove my skepticism wrong? I had to find out. Operator is an agent from OpenAI. Unlike most other agentic frameworks, which are designed to work with external APIs, Operator is designed to be fully autonomous with a web browser. More specifically, Operator is powered by a new model called Computer-Using Agent (CUA). It uses a combination of different models, including GPT-4o for vision to interact with graphical user interfaces. In practice, what this means is that you give it a goal, and on the Operator website, Operator will search the web to accomplish that goal for you. Pic: Operator building a list of financial influencers According to the OpenAI launch page, Operator is designed to ask for help (including inputting login details when applicable), seek confirmation on important tasks, and interact with the browser with vision (screenshots) and actions (typing on a keyboard and initiating mouse clicks). So, as soon as I gained access to Operator, I decided to give it a test run for a real-world task that any middle schooler can handle. Searching the web for influencers. Pic: A screenshot of the Operator webpage and the task I asked it to complete Why Do I Need Financial Influencers? For some context, I am building an AI platform to automate investing strategies and financial research. One of the unique features in the pipeline is monetized copy-trading. The idea with monetized copy trading is that select people can share their portfolios in exchange for a subscription fee. With this, both sides win – influencers can build a monetized audience more easily, and their followers can get insights from someone who is more of an expert. Right now, these influencers typically use Discord to share their signals and trades with their community. And I believe my platform can make their lives easier. Some challenges they face include: 1. They have to share their portfolios everyday manually, by posting screenshots. 2. Their followers have limited ways of verifying the influencer is trading how they claim they're trading. 3. Moreover, the followers have a hard time using the insights from the influencer to create their own investing strategies. Thus, with my platform NexusTrade, I can automate all of this for them, so that they can focus on producing content. Moreover, other features, like the ability to perform financial research or the ability to create, test, optimize, and deploy trading strategies, will likely make them even stronger investors. So these influencers win twice: one by having a better trading platform and again for having an easier time monetizing their audience. And so, I decided to use Operator to help me find some influencers. Giving Operator a Real-World Task I went to the Operator website and told it to do the following: Gather a list of 50 popular financial influencers from YouTube. Get their LinkedIn information (if possible), their emails, and a short summary of what their channel is about. Format the answers in a table Operator then opens a web browser and begins to perform the research fully autonomously with no prompting required. The first five minutes where extremely cool. I saw how it opened a web browser and went to Bing to search for financial influencers. It went to a few different pages and started gathering information. I was shocked. But after less than 10 minutes, the flaws started becoming apparent. I noticed how it struggled to find an online spreadsheet software to use. It tried Google Sheets and Excel, but they required signing in, and Operator didn't think to ask me if I wanted to do that. Once it did find a suitable platform, it began hallucinating like crazy. After 20 minutes, I told it to give up. If it were an intern, it would've been fired on the spot. Or if I was feeling nice, I would just withdraw its return offer.; tools: Operator. Human role: The original post does not say how a person stayed involved.; limitations: Operator requires a ChatGPT Pro subscription ($200/month). It may fail to proactively request human intervention for authentication steps, leading to workflow stalls. Hallucination risks are high during data aggregation tasks. Privacy and compliance concerns regarding automated scraping of personal contact information (emails, LinkedIn) were raised by community readers. The tool is currently exclusive to Pro subscribers..
Tools
Published on Agent Work Atlas:
Overview
I am the weirdest AI fanboy you'll ever meet.
Workflow
I am the weirdest AI fanboy you'll ever meet. I've used every single major large language model you can think of. I have completely replaced VSCode with Cursor for my IDE. And, I've had more subscriptions to AI tools than you even knew existed. This includes a $200/month ChatGPT Pro subscription. And yet, despite my love for artificial intelligence and large language models, I am the biggest skeptic when it comes to AI agents. Pic: "An AI Agent" — generated by X's DALL-E So today, when OpenAI announced Operator, exclusively available to ChatGPT Pro Subscribers, I knew I had to be the first to use it. Would OpenAI prove my skepticism wrong? I had to find out. Operator is an agent from OpenAI. Unlike most other agentic frameworks, which are designed to work with external APIs, Operator is designed to be fully autonomous with a web browser. More specifically, Operator is powered by a new model called Computer-Using Agent (CUA). It uses a combination of different models, including GPT-4o for vision to interact with graphical user interfaces. In practice, what this means is that you give it a goal, and on the Operator website, Operator will search the web to accomplish that goal for you. Pic: Operator building a list of financial influencers According to the OpenAI launch page, Operator is designed to ask for help (including inputting login details when applicable), seek confirmation on important tasks, and interact with the browser with vision (screenshots) and actions (typing on a keyboard and initiating mouse clicks). So, as soon as I gained access to Operator, I decided to give it a test run for a real-world task that any middle schooler can handle. Searching the web for influencers. Pic: A screenshot of the Operator webpage and the task I asked it to complete Why Do I Need Financial Influencers? For some context, I am building an AI platform to automate investing strategies and financial research. One of the unique features in the pipeline is monetized copy-trading. The idea with monetized copy trading is that select people can share their portfolios in exchange for a subscription fee. With this, both sides win – influencers can build a monetized audience more easily, and their followers can get insights from someone who is more of an expert. Right now, these influencers typically use Discord to share their signals and trades with their community. And I believe my platform can make their lives easier. Some challenges they face include: 1. They have to share their portfolios everyday manually, by posting screenshots. 2. Their followers have limited ways of verifying the influencer is trading how they claim they're trading. 3. Moreover, the followers have a hard time using the insights from the influencer to create their own investing strategies. Thus, with my platform NexusTrade, I can automate all of this for them, so that they can focus on producing content. Moreover, other features, like the ability to perform financial research or the ability to create, test, optimize, and deploy trading strategies, will likely make them even stronger investors. So these influencers win twice: one by having a better trading platform and again for having an easier time monetizing their audience. And so, I decided to use Operator to help me find some influencers. Giving Operator a Real-World Task I went to the Operator website and told it to do the following: Gather a list of 50 popular financial influencers from YouTube. Get their LinkedIn information (if possible), their emails, and a short summary of what their channel is about. Format the answers in a table Operator then opens a web browser and begins to perform the research fully autonomously with no prompting required. The first five minutes where extremely cool. I saw how it opened a web browser and went to Bing to search for financial influencers. It went to a few different pages and started gathering information. I was shocked. But after less than 10 minutes, the flaws started becoming apparent. I noticed how it struggled to find an online spreadsheet software to use. It tried Google Sheets and Excel, but they required signing in, and Operator didn't think to ask me if I wanted to do that. Once it did find a suitable platform, it began hallucinating like crazy. After 20 minutes, I told it to give up. If it were an intern, it would've been fired on the spot. Or if I was feeling nice, I would just withdraw its return offer.
Reported result
The execution failed to produce a reliable output. While the initial browsing was impressive, the agent struggled to find a suitable online spreadsheet tool, attempting Google Sheets and Excel without requesting necessary login credentials. Once it found a platform, it began hallucinating data. The user terminated the task after 20 minutes, comparing the performance to an intern who would have been fired.
Human role
The original post does not say how a person stayed involved.
Limitations
Operator requires a ChatGPT Pro subscription ($200/month). It may fail to proactively request human intervention for authentication steps, leading to workflow stalls. Hallucination risks are high during data aggregation tasks. Privacy and compliance concerns regarding automated scraping of personal contact information (emails, LinkedIn) were raised by community readers. The tool is currently exclusive to Pro subscribers.
Try it
I want to test OpenAI Operator for a lead generation task. Please access Operator and execute the following goal: 'Gather a list of 10 popular [replaceable: niche] influencers from [replaceable: platform]. For each, find their public email, LinkedIn URL, and a 2-sentence summary of their content. Format the output as a CSV table.' Steps: 1. Monitor the agent's browsing actions in real-time. 2. If the agent encounters a login wall or asks for credentials, provide them immediately or instruct it to skip that source. 3. If the agent starts hallucinating data or gets stuck on a tool selection (e.g., spreadsheet software), intervene by instructing it to output the data directly in the chat interface instead. 4. After the agent completes the task, verify the first 3 entries manually for accuracy. 5. Do not publish or use this data until I confirm the accuracy of the sample. If the agent fails to ask for login details when needed, note this as a limitation. If it hallucinates, stop the task and report the error. I will provide the specific niche and platform details before we start.
Source reports are not independently verified. Publication date is the catalog publication date, not the original article date.