Find London tours and order groceries
Hands on with Operator — a promising but frustrating new frontier for artificial intelligence. Task: Hands on with Operator — a promising but frustrating new frontier for artificial intelligence This is a column about AI. My boyfriend works at Anthropic. See my full ethics disclosure here. For the past few months, the big artificial intelligence labs have signaled that 2025 will be a breakout year for agents: tools that use computers on your behalf, working across different apps to perform multi-step operations just like humans do. Agents have rapidly become a top priority at AI labs, as they represent a key milestone on the way toward building full-featured virtual coworkers. Starting last year, a steady drip of announcements signaled that researchers are rapidly making progress toward that goal. Anthropic introduced its first take on agents, which it calls “computer use,” into its API in October. Two months later, Google said that its Gemini 2.0 models were designed for “the agentic era.” Despite those announcements, though, until now most agents have only been available to developers. As of this week, though, that’s starting to change. On Wednesday, Google announced that the mobile version of Gemini can now carry out tasks across different apps, such as searching for upcoming sports games and creating calendar entries for them, with a single prompt. A day later, the upstart AI company Perplexity announced a similar agent for its Android app. And in the most ambitious AI agent launch to date, on Thursday OpenAI announced Operator: an AI agent that can browse the web and take actions on your behalf. In a live demo on Thursday morning, CEO Sam Altman and three of his coworkers showed some of what Operator can do: shop for groceries on Instacart, hunt for concert tickets on Stubhub, and booking a reservation at a local restaurant through OpenTable. Operator is available through what OpenAI is calling a “research preview,” and is currently available only to users of the $200-a-month ChatGPT Pro tier. The company plans to release it to users of its $20-a-month Plus plan and other users in the months ahead, it said. For developers, it will be available in the company’s API within a few weeks, the company said. OpenAI explained how Operator works in a blog post: After OpenAI’s demo, I paid to upgrade my ChatGPT subscription so I could use Operator. A few hours later, it was enabled on my account, and I spent the afternoon trying various prompts to get a sense of what it can do. How you feel about Operator will depend heavily on what you hope to get out of it. If you’re looking for a polished virtual assistant to which you can confidently hand off shopping and research tasks, I suspect you’ll be disappointed: at the moment, using Operator is significantly slower, more frustrating, and more expensive than simply doing any of these tasks yourself. If you’re interested in the bleeding edge of AI progress, though, and want to get a sense of what the future could look like, Operator offers a compelling demonstration. Using it on Thursday reminded me a bit of the first time I took a ride in one of Google’s self-driving cars: half-baked though the product may have been at the time, it also represented an extraordinary technological achievement and seemed likely to shape our world for years to come. Of all the tasks I gave to Operator, the one it performed best was probably the simplest. (It was also a task that Operator suggested I try in a grid of suggested prompts under the chatbot interface.) The task was to “suggest some top-rated walking tours I can sign up for in London.” Once I hit “go,” Operator opened TripAdvisor in a browser, did a search, and presented me with a series of walking tours to consider with the relevant links. The whole thing only took about a minute, and I could have easily left the browser tab to do something else while it worked. Of course, there’s nothing that impressive about generating a list of London walking tours: there are many such lists all over the web, and the free version of ChatGPT will also give you a good list even faster than Operator. But there’s still something striking about telling a computer to do something and then watch it open a browser tab, type queries, click on buttons and present you with a report.; tools: Operator. Human role: The original post does not say how a person stayed involved.; limitations: Operator uses a separate browser instance, meaning it lacks existing user sessions, cookies, or saved preferences, leading to repeated logins and location errors. It cannot access sites that block OpenAI crawlers (e.g., YouTube, Reddit). The tool is currently a research preview limited to ChatGPT Pro subscribers ($200/month). Human oversight is critical for authentication and correcting agent mistakes..
Tools
Published on Agent Work Atlas:
Overview
Hands on with Operator — a promising but frustrating new frontier for artificial intelligence
Workflow
Hands on with Operator — a promising but frustrating new frontier for artificial intelligence This is a column about AI. My boyfriend works at Anthropic. See my full ethics disclosure here. For the past few months, the big artificial intelligence labs have signaled that 2025 will be a breakout year for agents: tools that use computers on your behalf, working across different apps to perform multi-step operations just like humans do. Agents have rapidly become a top priority at AI labs, as they represent a key milestone on the way toward building full-featured virtual coworkers. Starting last year, a steady drip of announcements signaled that researchers are rapidly making progress toward that goal. Anthropic introduced its first take on agents, which it calls “computer use,” into its API in October. Two months later, Google said that its Gemini 2.0 models were designed for “the agentic era.” Despite those announcements, though, until now most agents have only been available to developers. As of this week, though, that’s starting to change. On Wednesday, Google announced that the mobile version of Gemini can now carry out tasks across different apps, such as searching for upcoming sports games and creating calendar entries for them, with a single prompt. A day later, the upstart AI company Perplexity announced a similar agent for its Android app. And in the most ambitious AI agent launch to date, on Thursday OpenAI announced Operator: an AI agent that can browse the web and take actions on your behalf. In a live demo on Thursday morning, CEO Sam Altman and three of his coworkers showed some of what Operator can do: shop for groceries on Instacart, hunt for concert tickets on Stubhub, and booking a reservation at a local restaurant through OpenTable. Operator is available through what OpenAI is calling a “research preview,” and is currently available only to users of the $200-a-month ChatGPT Pro tier. The company plans to release it to users of its $20-a-month Plus plan and other users in the months ahead, it said. For developers, it will be available in the company’s API within a few weeks, the company said. OpenAI explained how Operator works in a blog post: After OpenAI’s demo, I paid to upgrade my ChatGPT subscription so I could use Operator. A few hours later, it was enabled on my account, and I spent the afternoon trying various prompts to get a sense of what it can do. How you feel about Operator will depend heavily on what you hope to get out of it. If you’re looking for a polished virtual assistant to which you can confidently hand off shopping and research tasks, I suspect you’ll be disappointed: at the moment, using Operator is significantly slower, more frustrating, and more expensive than simply doing any of these tasks yourself. If you’re interested in the bleeding edge of AI progress, though, and want to get a sense of what the future could look like, Operator offers a compelling demonstration. Using it on Thursday reminded me a bit of the first time I took a ride in one of Google’s self-driving cars: half-baked though the product may have been at the time, it also represented an extraordinary technological achievement and seemed likely to shape our world for years to come. Of all the tasks I gave to Operator, the one it performed best was probably the simplest. (It was also a task that Operator suggested I try in a grid of suggested prompts under the chatbot interface.) The task was to “suggest some top-rated walking tours I can sign up for in London.” Once I hit “go,” Operator opened TripAdvisor in a browser, did a search, and presented me with a series of walking tours to consider with the relevant links. The whole thing only took about a minute, and I could have easily left the browser tab to do something else while it worked. Of course, there’s nothing that impressive about generating a list of London walking tours: there are many such lists all over the web, and the free version of ChatGPT will also give you a good list even faster than Operator. But there’s still something striking about telling a computer to do something and then watch it open a browser tab, type queries, click on buttons and present you with a report.
Reported result
The author reports mixed outcomes: Operator successfully found London tours in about a minute but failed to efficiently order groceries, requiring 15 minutes of manual intervention due to location and login issues. The agent is described as slower and more frustrating than doing tasks manually. No quantitative business metrics are provided, only qualitative assessments of speed and reliability.
Human role
The original post does not say how a person stayed involved.
Limitations
Operator uses a separate browser instance, meaning it lacks existing user sessions, cookies, or saved preferences, leading to repeated logins and location errors. It cannot access sites that block OpenAI crawlers (e.g., YouTube, Reddit). The tool is currently a research preview limited to ChatGPT Pro subscribers ($200/month). Human oversight is critical for authentication and correcting agent mistakes.
Try it
Act as an AI agent using OpenAI Operator to research [replaceable input: e.g., top-rated walking tours in London] and attempt to [replaceable input: e.g., add specific grocery items to an Instacart cart]. First, execute a read-only search task and present the results with links for my review. Do not proceed to any transactional step (like adding to cart or booking) without my explicit confirmation. If you encounter a login screen, location mismatch, or CAPTCHA, stop immediately and hand control back to me. For the grocery task, verify the delivery location is [replaceable input: e.g., San Francisco] before searching. Measure success by whether the correct items are identified without manual correction. If the agent fails to locate the correct store or requires more than 3 steps to resolve a location error, abort the task and report the failure. I will manually complete any transactional steps after reviewing your proposed actions.
Source reports are not independently verified. Publication date is the catalog publication date, not the original article date.