OpenAI announced the launch of ChatGPT Agent, a proxy function with general operational capabilities, allowing AI services to transform from simply answering questions to an agent-type tool that can truly "complete tasks" on behalf of users.
Currently, this new feature is available to ChatGPTPro, Plus and Team subscribers, and may be provided to general free users in a limited form in the future.
This newly launched agent has the ability to automatically browse calendars, generate presentations, write slides, and even run code directly on the built-in terminal. It also integrates several important recent technological achievements of OpenAI, including the recently launched ability to click and interact with web pages.Operator, and Deep Research, which can compile data and produce complete reports.
ChatGPT can now do work for you using its own computer.
Introducing ChatGPT agent—a unified agentic system combining Operator's action-taking remote browser, deep research's web synthesis, and ChatGPT's conversational strengths. pic.twitter.com/7uN2Nc6nBQ
- OpenAI (@OpenAI) July 17, 2025
Natural language input can perform multiple complex tasks
Usage is incredibly simple. Simply enable Agent mode in the ChatGPT tool menu to access the ChatGPT Agent interface. By entering commands through natural language, users can trigger highly complex actions such as "planning a trip to the US," "procuring ingredients for a Japanese breakfast," or "analyzing three competitors and creating a presentation." It can also integrate with various connectors, connecting to services like Gmail and GitHub, or calling external applications through APIs to execute more advanced workflows.
In this wave of updates, OpenAI announced that it will retire the Operator function in 30 days, but the Deep Research module will still be retained as an independent option to provide scenario applications that require more in-depth research.
Test results show a significant lead
According to test data released by OpenAI, ChatGPT Agent achieved impressive performance in multiple AI tests, including a 1% pass on Humanity's Last Exam (pass@41.6), roughly double the performance of o3 and o4-mini. FrontierMath achieved a 2% score, significantly higher than o27.4-mini's 4%. This demonstrates that ChatGPT Agent is more than just an AI language model; it's a multi-tasking AI tool with both reasoning and execution capabilities.
Attach great importance to safety and user control
As agent-based AI tools gradually enter the mainstream market, OpenAI has also made it clear that its ChatGPT Agent agent function attaches great importance to security control in its design. Before anything else, such as booking a trip, filling out a form, or interacting with a website involving personal information, users will be asked for explicit authorization.
OpenAI has also introduced a "supervised mode," requiring users to incrementally review each action during high-risk tasks. When faced with sensitive content like financial transactions and legal advice, the ChatGPT Agent function will proactively refuse to execute.
Universal Agent is an important step towards the vision of AI work assistant
While numerous AI agent tools are currently available, most remain insufficient for complex tasks. OpenAI emphasizes that ChatGPT Agent, powered by its robust foundational model and mature integration framework, is the most competitive general-purpose agent-based AI tool on the market. It elevates ChatGPT's application from a reactive approach to true "executive AI."
