Openai CEO Sam Altman kicked off this year by saying in a blog post that 2025 is big for AI agents. This is a tool that can automate tasks and take actions for you.
Now we’re looking at Openai’s first real attempt.
Openai on Thursday announced the launch of a research preview of Operator, a general-purpose AI agent that can take control of a web browser and perform certain actions independently. Operators are coming to users with CHATGPT’s $200 Pro subscription plan. Openai says it plans to roll out this feature to more users in PLUS, Teams, and Enterprise stages.
“(Operators) will soon be in other countries,” Openai CEO Sam Altman said in a livestream Thursday. “Unfortunately, Europe will take some time.”
This initial research preview is available at operator.chatgpt.com, but soon, Openai says it wants to integrate Operator into all CHATGPT clients.
According to Openai, the operator promises to automate tasks such as booking travel accommodations, restaurant reservations, and online shopping. There are several task categories that you can select from within the operator interface. This includes shopping, delivery, dining, travel, and more. All of these enable different types of automation.
When a ChatGPT user activates an operator, a small window pops up displaying a dedicated web browser that the agent uses to complete tasks, and a description of the specific action the agent is performing. The operator uses its own dedicated browser, so you can control the screen while the operator is running.
Openai says the operator is powered by a computer-assisted agent model (CUA) that combines the vision capabilities of the company’s GPT-4O model with the inference capabilities of Openai’s more advanced models. CUA is trained to interact with a website’s front end. This means you don’t need to use developer APIs to leverage various services.
In other words, CUAs can use buttons, navigate menus, and fill out forms on web pages just like humans.
Openai says it works with companies like Doordash, eBay, Instacart, Priceline, StubHub, and Uber to ensure operators honor these companies’ terms of service agreements.
“CUA models are trained to request user confirmation before submitting an order before finalizing a task with external side effects. This allows the model to ,” Openai wrote in a document provided to TechCrunch. “(It) has already proven useful in a variety of cases, and we aim to extend its reliability across a broader range of tasks.”
However, Openai warns that CUA is not perfect. The company said it “does not yet expect CUA to work reliably in all scenarios.”
“Currently, operators cannot reliably handle many complex or specialized tasks,” Openai adds in its support documentation.
Among the abundance of attention, Openai also requires supervision of some tasks such as banking transactions, CUA and operators. For example, the user must take over to enter credit card information. Openai says its operators do not collect data or screenshot.
“For particularly sensitive websites, such as email, operators require active user supervision so that users can directly catch and address mistakes that the model may make,” Openai said in support materials. I am.
This certainly limits the usefulness of the operator, but also ensures that the agent does not interfere with the illusion and, for example, spends the mortgage payment on an accent chair. Google took a similar approach with its Project Mariner AI agent, which also doesn’t fill in information like credit card numbers.
limit
Operators have some limitations worth noting.
There are rate limits – both daily and task dependent. Openai says operators can perform multiple tasks at once, but there are “dynamic limits” to this. There is also an overall usage limit that resets daily.
At this release stage, operators also refuse to perform tasks entirely for security reasons, such as sending emails (the fact that CUA allows this) and deleting calendar events. Openai says this will change in the future, but won’t give an ETA.
It may also become “stuck” if the operator runs into a particularly complex interface, password field, or Captcha check. According to Openai, they will ask the user to take over when this happens.
The future of agents
Openai has been fairly slow in developing its AI agents compared to its rivals (see Rabbit, Google, and Humanity’s agents), which may be associated with safety risks regarding the technology.
If AI systems can perform actions on the web, it opens the door to far more dangerous use cases from malicious actors. You can automate AI agents to orchestrate phishing scams, DDO attacks, or grab concert tickets before someone else can. It is important that Openai takes steps to prevent these types of exploitation, especially in the case of ChatGpt and widely used tools.
Openai seems to think it’s safe enough to release in its current form, at least as a research preview.
“Operators employ tools that attempt to limit the susceptibility of models to malicious prompts, hidden instructions, and phishing attempts,” Openai explains on its website. “Monitoring systems pause execution if suspicious activity is detected, while automated, human-reviewed pipelines continually update safeguards.”
Operator is Openai’s boldest attempt at creating an AI agent. Last week, Openai released Tasks, which provides simple automation capabilities for ChatGpt, giving you the ability to set reminders and schedule prompts to run at set times each day.
The task provided ChatGpt users with familiar but necessary features that make ChatGpt practical to use as Siri or Alexa. However, operators have shown capabilities that previous generations of virtual assistants cannot.
AI agents are being touted as the next big thing in AI after CHATGPT. This is a new technology that will change the way people use the Internet and their PCs. Instead of simply distributing and processing information, agents can theoretically take action and actually do things.
With the release of Openai’s first concrete agent, it quickly becomes clear how realistic this vision is.