ChatGPT’s new AI agent can browse the web and create PowerPoint slideshows

On Thursday, OpenAI introduced an innovative feature called ChatGPT Agent, designed to enhance the capabilities of its AI assistant by enabling it to perform complex, multi-step tasks through its own web browser. This new functionality combines elements from OpenAI's previous tools, Operator and Deep Research, allowing ChatGPT to navigate websites, execute code, and generate documents while keeping users in control of the process. This advancement represents OpenAI's foray into 'agentic AI'—intelligent systems capable of executing autonomous actions for users. With ChatGPT Agent, users can delegate various tasks, such as assembling outfits for specific events, crafting PowerPoint presentations, meal planning, or updating financial spreadsheets with fresh data. The system employs a blend of web browser usage, terminal access, and API connections, including ChatGPT Connectors that work with applications like Gmail and GitHub. While utilizing the Agent, users can observe a window within the ChatGPT interface that displays the AI's activities in a secure environment. This sandbox operates independently with its own virtual operating system and web browser, ensuring it does not interfere with the user's personal device. OpenAI explains, "ChatGPT executes these tasks using its virtual computer, seamlessly transitioning between reasoning and action to manage intricate workflows based on your directives." Similar to the previous Operator feature, the Agent requires user consent before undertaking tasks that have real-world implications, like making purchases. Users also have the ability to halt tasks at any moment, regain control of the browser, or cease operations entirely. Additionally, a "Watch Mode" is available for tasks requiring direct user oversight, such as sending emails. Although the capabilities of ChatGPT Agent extend beyond those of its predecessor, OpenAI has announced that the Operator preview site will remain operational for a few more weeks before its discontinuation. However, the effectiveness of this new AI agent in executing multi-step tasks will likely vary depending on the context. It is essential to note that the AI model, while sophisticated, is not an all-encompassing problem-solving entity, but rather a proficient imitator. Its ability to adapt to new scenarios is limited; tasks that fall outside its training examples may still pose challenges.