According to the official announcement of OpenAI and the technical live broadcast of July 17th, ChatGPT Agent is a unified intelligence body that combines Operator ‘ s web-based interactive capabilities, Deep Research ‘ s information integration and ChatGPT ‘ s dialogue intelligence with self-thinking and action skills. Users can use natural language tips to allow ChatGPT Agent to perform multi-step complex tasks on its own in a virtual computer environment. Functions include browsing web pages, running codes, generating editable slides and spreadsheets.
ChatGPT Agent was rolled out to Pro, Plus and Team users on 17 July (400 messages per month for Pro users, 40 for Plus and Team users), and business and education users will be granted access in the coming weeks, with the European Economic Area and Switzerland not open. ChatGPT Agent inherited Operator’s web-based interactive function by clicking, rolling and entering text through virtual browsers, and processing content requiring user authentication (e.g. login to Gmail or GitHub). By integrating Deep Research ‘ s multi-step information integration, hundreds of online sources (e.g. text, images, PDF) can be analysed to generate detailed studies better than the separate Deep Research model.
In combination with ChatGPT ‘ s fluidity of dialogue and reasoning, Agent can develop action plans, dynamically adjust strategies and generate deliverables (e.g. PowerPoint or Excel documents). OpenAI states that it performed excellently in a number of benchmark tests, such as Humanity’s Last Exam (41.6 per cent single pass rate) and Frontier Math (27.4 per cent accuracy rate, which exceeds 6.3 per cent for o4-mini). OpenAI has deployed strict security mechanisms for ChatGPT Agent, including real-time monitoring, high-risk biological and chemical classification, the prohibition of memory functions to prevent misuse, and restrictions on high-risk tasks such as financial transactions. Users are required to authorize sensitive operations and control Agent’s behaviour through Watch Mode or Takeover Mode.
OpenAI plans to continue the iterative version of ChatGPT Agent, to enhance its ability to handle complex workflows and possibly integrate them with the anecdotal GPS-5 to achieve more seamless voice, visual and mission performance. Current constraints include slow processing and reasoning limitations for new and complex types of missions. In the future, OpenAI may introduce independent browsers or further optimize API access. ChatGPT Agent is the result of OpenAI ‘ s in-house research and development that integrates a variety of AI capabilities (web-based interaction, research, dialogue) to achieve autonomous mission management. OpenAI promotes “agent AI” development through ChatGPT Agent, with the goal of transforming AI from a tool of dialogue to an autonomous productivity partner, challenging competitors such as Anthropic Claude.
