- OpenAI is rumored to be launching new AI agent tool in January
- Report claims it's codenamed Operator and can control your browser
- This help could help it write code or book travel on your behalf
Forget chatbots – AI's next big trick is expected to be artificial intelligence agents that can carry out computer-based chores on your behalf, and a new report claims OpenAI's version will be landing very soon.
OpenAI's new tool, codenamed 'Operator,' could be released in January, according to Bloomberg and its two internal sources. It will apparently be released as a research preview initially and through the company's application programming interface (API) for developers to get their hands on.
AI agents are designed to operate your computer on your behalf without supervision. Like Anthropic's similar 'computer use' tool for Claude, OpenAI's Operator is expected to be able to carry out actions like writing code or booking travel.
The space is shaping up to be the next big AI battleground, with Google recently leaking its so-called Jarvis AI tool that can browse the web for you. According to The Information, it could beat OpenAI to the punch by landing in December.
Like Google's solution, OpenAI's equivalent is expected to be a "general purpose tool that executes tasks in a web browser," according to Bloomberg's source. So, while it might not be quite ready to pilot desktop apps on your behalf, OpenAI's Operator sounds like another pretty powerful browser extension.
How much do you trust AI?
There's no doubt that tech giants think AI agents are the next big computing platform – in a recent Reddit AMA, OpenAI CEO Sam Altman said, "I think the thing that will feel like the next giant breakthrough will be agents."
This apparent breakthrough will come soon, too, with Altman adding, “IMHO this is going to be a big theme in 2025.” Recent announcements and leaks from Anthropic, Microsoft, Google, and OpenAI certainly support this.
However, given that AI agents are expected to work autonomously without supervision, they could also demand a large amount of trust. A significant asterisk on tools like ChatGPT Search – which already searches the web for you – is that they're early releases that can still make mistakes.
In October, OpenAI released a new benchmark called 'SimpleQA' to measure the accuracy of its own AI models – and its recent o1 preview model performed pretty poorly, only getting a 42% success rate in the new benchmark.
So, while AI agents might be the next big hype feature in 2025, it could still be a while before we entrust them with automating our computer-based chores – at least without some careful supervision.