Meet OpenAI’s new web browsing agent.
On Thursday, Openai announced the Chatgpt Agent, a tool that is able to navigate the network and perform tasks on your behalf. When teased in X posts before the live broadcast, Chatgpt Agent combines the autonomy of its operators with the inference intelligence of its deep research tools.
OpenAI’s carrier launched ChatGpt Pro users in preview mode in January, which can scroll, click and enter online, but has limitations but has never seen widespread releases. Deep research is another type of agent that can search the network and compile information on behalf of users, but it cannot take action. OpenAI’s new web browsing agent launch effectively combines these two tools.
Image source: OpenAI
I’m testing the confused new AI browser Comet. Here are 3 features I tried immediately.
“Operators cannot delve into the analysis or write detailed reports, and in-depth research cannot interact with the website to refine the results or access the user authentication access,” OpenAI announcement said. “We see many questions with the operators trying to use are actually more suitable for in-depth research, so we put both together.”
Openai’s new tools are part of the fast-growing world of proxy AI
OpenAI’s ChatGpt proxy joins other recently released proxy tools that can perform tasks on behalf of users. Although not a complete web browser, it works similar to the browser assistant of Pyplexity Comet. Anthropic also has a tool called “Computer Use” that can take over your cursor and write code. As models become more advanced, they are more capable of performing autonomous tasks. Web browsing is considered one of the next areas of AI labs, and can compete for OpenAI, humans, confusing already shipped features, and Google’s Project Mariner Research Prototype.
Mixable light speed
An OpenAi spokesperson said that when you are in Agent Mode, you can ask Chatgpt to do “View my calendar and briefly introduce me to my client meetings based on recent news” or “plan and buy ingredients to make Japanese breakfast four”.
Openai said that users are always in control and requesting permission before taking measures, such as submitting forms, making purchases or processing personal information. Openai says users can easily interrupt or take over if needed. Some tasks, such as sending emails, require supervision, called watch mode.
On the security side, Chatgpt agents are not allowed to perform “high-risk” tasks, such as financial transactions or legal advice, and are trained to identify malicious or confrontational attacks from rapid injection or other manipulation strategies.
In terms of data collection, OpenAI says Chatgpt can delete its browsing data with one click and log out of the website, and will not collect data when users are actively participating in tasks such as entering passwords.
Given the progress, the risk of problems being wrong is also higher. Openai classifies Chatgpt agents as their high risk in the preparation framework for biological and chemical abilities.
This is good news for users who don’t have a $200 monthly ChatGpt Pro plan. It is available for Pro, Plus and Team users. PRO users will be accessible today, and team users will be given access in the next few days. Pro users receive 400 messages per month, while Plus, Team and other paid users receive 40 messages per month.
Disclosure: Mashable’s parent company Ziff Davis filed a lawsuit against Openai in April, accusing it of infringing on Ziff Davis’ copyright in training and operating its AI systems.
