OpenAI is going all-in on the most-hyped trend in AI right now: AI agents, or tools that go a step beyond chatbots to complete complex, multi-step tasks on a user\u2019s behalf. The company on Thursday debuted ChatGPT Agent, which it bills as a tool that can complete work on your behalf using its own \u201cvirtual computer.\u201d<\/p>\n
In a briefing and demo with The Verge, Yash Kumar and Isa Fulford \u2014 product lead and research lead on ChatGPT Agent, respectively \u2014 said it\u2019s powered by a new model that OpenAI developed specifically for the product. The company said the new tool can perform tasks like looking at a user\u2019s calendar to brief them on upcoming client meetings, planning and purchasing ingredients to make a family breakfast, and creating a slide deck based on its analysis of competing companies.<\/p>\n
The model behind ChatGPT Agent, which has no specific name, was trained on complex tasks that require multiple tools \u2014 like a text browser, visual browser, and terminal where users can import their own data \u2014 via reinforcement learning, the same technique used for all of OpenAI\u2019s reasoning models. OpenAI said that ChatGPT Agent combines the capabilities of both Operator and Deep Research, two of its existing AI tools.<\/p>\n
To develop the new tool, the company combined the teams behind both Operator and Deep Research into one unified team. Kumar and Fulford told The Verge that the new team is made up of between 20 and 35 people across product and research.<\/p>\n
In the demo, Kumar and Fulford demonstrated potential use cases for ChatGPT Agent, like asking it to plan a date night by connecting to Google Calendar to see when the user has a free evening, and then cross-referencing OpenTable to find openings at certain types of restaurants. They also showed how a user could interrupt the process by adding, say, another restaurant category to search for. Another demonstration showed how ChatGPT Agent could generate a research report on the rise of Labubus versus Beanie Babies.<\/p>\n
Fulford said she enjoyed using it for online shopping because the combination of tech behind Deep Research and Operator worked better and was more thorough than trying the process solely using Operator. And Kumar said he had begun using ChatGPT Agent to automate small parts of his life, like requesting new office parking at OpenAI every Thursday instead of showing up Monday having forgotten to request it with nowhere to park.<\/p>\n
Kumar said that since ChatGPT Agent has access to \u201can entire computer\u201d instead of just a browser, they\u2019ve \u201cenhanced the toolset quite a bit.\u201d<\/p>\n
According to the demo, though, the tool can be a bit slow. When asked about latency, Kumar said their team is more focused on \u201coptimizing for hard tasks\u201d and that users aren\u2019t meant to sit and watch ChatGPT Agent work.<\/p>\n
\u201cEven if it takes 15 minutes, half an hour, it\u2019s quite a big speed-up compared to how long it would take you to do it,\u201d Fulford said, adding that OpenAI\u2019s search team is more focused on low-latency use cases. \u201cIt\u2019s one of those things where you can kick something off in the background and then come back to it.\u201d<\/p>\n
Before ChatGPT Agent does anything \u201cirreversible,\u201d like sending an email or making a booking, it asks for permission first, Fulford said.<\/p>\n
Since the model behind the tool has increased capabilities, OpenAI said it has activated the safeguards it created for \u201chigh biological and chemical capabilities,\u201d even though the company said it does not have \u201cdirect evidence that the model could meaningfully help a novice create severe biological or chemical harm\u201d in the form of weapons. Anthropic in May activated similar safeguards<\/a> for its launch of one of its Claude models, Opus 4.<\/p>\n When asked about whether the tool is permitted to perform financial transactions, Kumar said those actions have been restricted \u201cfor now,\u201d and that there\u2019s an additional protection called Watch Mode, wherein if a user navigates to a certain category of webpages, like financial sites, they must not navigate away from the tab ChatGPT Agent is operating in or the tool will stop working.<\/p>\n OpenAI will start rolling out the tool today to Pro, Plus, and Team users \u2014 pick \u201cagent mode\u201d in the tools menu or type \u201c\/agent\u201d to access it \u2014 and the company said it will make it available to ChatGPT Enterprise and Education users later this summer. There\u2019s no rollout timeline yet for the European Economic Area and Switzerland.<\/p>\n The concept of AI agents has been a buzzworthy trend in the industry for years. The ideal developers are working toward is something like Iron Man\u2019s J.A.R.V.I.S., a tool that can perform specific job functions, check people\u2019s calendars for the best time to schedule an event, purchase a gift based on a friend\u2019s preferences, and more, but at the moment, they\u2019re somewhat limited to assisting with coding and compiling research reports.<\/p>\n The term \u201cAI agent\u201d became more common to investors and tech executives in 2023 and quickly picked up speed, especially after fintech company Klarna announced in February 2024 that in just one month of operation, its own AI agent had handled two-thirds of its customer service chats \u2014 the equivalent of 700 full-time human workers. From there, executives at Amazon, Meta, Google, and more started mentioning their AI agent goals on earnings call after earnings call<\/a>. And since then, AI companies have been strategically hiring to reach those goals: Google, for instance, last week<\/a> hired Windsurf\u2019s CEO, co-founder and some R&D team members to help further its agentic AI projects.<\/p>\n OpenAI\u2019s debut of ChatGPT Agent follows its January release of Operator, which the company billed as<\/a> \u201can agent that can go to the web to perform tasks for you\u201d since it was trained to be able to handle the internet\u2019s buttons, text fields and more. It\u2019s also part of a larger trend in AI, as companies large and small chase AI agents that will capture the attention of consumers and ideally become habits. Last October, Anthropic, the Amazon-backed AI startup behind Claude, released a similar tool called \u201cComputer Use,\u201d which it billed as a tool that could use a computer the same way a human can in order to complete tasks on a user\u2019s behalf. Multiple AI companies, including OpenAI, Google and Perplexity, also offer an AI tool that all three have dubbed Deep Research, denoting an AI agent that can write sizable analyses and research reports on anything a user wants.<\/p>\n