OpenAI on Thursday, July 17, 2025 announced ChatGPT Agent, a new agentic capability for ChatGPT that can autonomously carry out complex, multi-step tasks on the user’s behalf.
ChatGPT Agent “thinks and acts” using its own virtual computer — including a browser, code console, and app connectors — to browse websites, run code, and produce finished outputs (like slides or spreadsheets).
In practice, users can now ask ChatGPT Agent to do things like “look at my calendar and brief me on upcoming client meetings based on recent news,” “plan and buy ingredients to make Japanese breakfast for four,” or “analyze three competitors and create a slide deck”.
These tasks are carried out end-to-end, with ChatGPT Agent navigating sites, logging in securely, and returning editable results while keeping the user in control.
Key Features
- Unified Agentic System: ChatGPT Agent merges the capabilities of OpenAI’s previous tools. It combines Operator (web browsing/interaction) and Deep Research (information synthesis) within one model, along with ChatGPT’s conversational AI. The result is a single agent that can research and act across the web in a unified workflow.
- Multi-Tool Virtual Environment: The agent runs on its own “virtual computer” that includes a visual web browser, a text-based browser, a Linux-like terminal, and direct API access. It can switch tools as needed – for example, using the text browser to skim long documents and the GUI browser to click through sites. Crucially, ChatGPT Agent can “take over” the browser and log in to sites (preserving user privacy) to access personal apps and data.
- App Connectors: Users can link external services via ChatGPT Connectors (e.g. Gmail, Google Calendar, Google Drive, GitHub, etc.). Once granted access, the agent can pull in real data — for instance, summarizing your email inbox or finding free time slots in your calendar — before asking for login confirmation to act on them.
- Rich Outputs: ChatGPT Agent can produce fully editable files. It can generate PowerPoint slides or Excel spreadsheets populated with data, charts, and text that the user can tweak later. OpenAI’s tests even show it outperforming prior tools on benchmarks: for example, on one spreadsheet-editing task the agent scored 45.5% accuracy versus 20.0% for Microsoft’s Copilot in Excel. Downloaded web content or data can be manipulated via the terminal before final output.
- Interactive Workflow: The system is designed to be collaborative and interruptible. As the agent works, users can pause or steer the process at any time. The agent may ask follow-up questions to clarify goals, or the user can interject with new instructions. If a task is long-running, users can request progress updates or let the agent run in the background and notify them upon completion (the mobile ChatGPT app will alert you when it’s done).
Use Cases
ChatGPT Agent greatly expands what users can automate. Possible examples include:
Slide Decks and Reports:
Ask the agent to research a topic (e.g. “summarize competitor strategies”) and it will gather information from the web and deliver an editable presentation. It can create slideshows with text, charts and images that you can further refine. (In demos, the agent was shown building a slide deck on Nvidia’s earnings report and on-the-fly competitor analyses.)
Spreadsheets and Data:
The agent can update or build spreadsheets with formulas, charts, and proper formatting. For instance, it could pull the latest financial data and insert it into a budget model or amortization schedule. OpenAI says the agent surpasses previous models on data-science and spreadsheet tasks.
Research and Analysis:
It can surf the web like an employee researcher: running multi-site searches, compiling notes, summarizing long documents, or coding quick analyses. For example, it might “analyze three competitors and create a slide deck” or write a concise report on a technical subject by reading multiple papers.
Shopping and Planning:
The agent can handle online shopping and event planning. Demo examples include planning a family meal by finding recipes and ordering ingredients, or even ordering an outfit (accounting for weather and dress code). Wired reporter Isa Fulford gave the agent a cumbersome cupcake order; it took about an hour, but successfully placed the order online, a task she found easier to delegate.
Email and Calendar:
By connecting to your Google or Outlook accounts, ChatGPT Agent can summarize your inbox (e.g. highlight important emails or action items) and manage your calendar. It might brief you on upcoming meetings based on recent messages, or suggest and book times for new appointments. It still requires the user to confirm any email sends or scheduling changes, but it can do the heavy lifting of identifying times and drafting content.
Task Automation:
Routine workflows can be automated. For example, the agent could convert a web dashboard into a slide presentation, rearrange meetings based on conflicts, or plan and book a business offsite. The Verge noted an OpenAI employee using the agent to automate a weekly parking request. Anything that involves clicking through websites, filling forms, or combining data from multiple sources is on the table.
Launch and Availability
OpenAI is rolling out ChatGPT Agent starting July 17, 2025. Pro, Plus, and Team subscribers will be the first to get access. Pro users receive immediate access, while Plus and Team users should see it over the next few days.
Enterprise and Education customers will be onboarded in the following weeks. (OpenAI notes that access is still being enabled for Europe).
At launch, each Pro subscriber can run up to 400 agent tasks per month; other paid tiers get 40 per month, with additional usage available via credits. The agent is enabled by choosing “agent mode” from ChatGPT’s tools menu during any chat.
Safety and User Controls
OpenAI emphasizes that the user is always in control of the agent. The system is built with multiple safeguards for risky actions. Before executing anything with real-world impact (like buying something or sending an email), ChatGPT Agent explicitly asks for your permission.
Sensitive tasks (such as sending messages or making bookings) are subject to an optional “watch mode” requiring you to actively supervise or approve each step. OpenAI has also hard-coded refusals for extremely high-risk actions (e.g. bank transfers).
To prevent malicious manipulation, the agent is trained to recognize and reject “prompt injection” attempts (hidden instructions on web pages). In practice, this means the agent should ignore hidden malicious instructions in websites it visits, and it will not act on any request without asking you first.
Privacy protections include a special browser takeover mode: when you let the agent control a browser, your inputs (like passwords) remain private and are not sent to OpenAI.
You can also delete all browsing data and logout of sessions with one click, and the agent does not store any personal data after the task completes.
In remarks on the launch, OpenAI CEO Sam Altman stressed that ChatGPT Agent is cutting-edge and experimental. He advised users to treat it as a “chance to try the future” but cautioned against using it for high-stakes or highly personal tasks until it has been proven safe in the wild.
This aligns with OpenAI’s official stance: they describe ChatGPT Agent as a powerful new tool that expands real-world utility, but one that must be used judiciously.
Outlook
OpenAI presents ChatGPT Agent as “bridging research and action” – a step toward AI that not only informs but also executes tasks. The company says this launch is just the beginning, with continuous improvements planned to make the agent faster and more capable over time.
For users eager to automate work and personal tasks, ChatGPT Agent promises a new level of AI assistance. However, both OpenAI and independent observers emphasize that it must be used with care: it’s powerful, but still learning and restricted by safety guards.
As ChatGPT Agent rolls out to more subscribers this summer, it will be closely watched as a bellwether for the future of AI-powered work automation.