Introduction to ChatGPT agent

ChatGPT agent is positioned as a unified “do-the-work” system that can plan, browse, and act across a long task horizon—using a virtual computer, a text browser, a visual (GUI) browser, and a terminal—so users can go from ideas to completed outcomes like bookings, documents, and shopping. The core shift is consolidation: instead of choosing between separate capabilities for deep web research and interactive web actions, the agent switches among tools as needed, runs code, calls APIs, and produces artifacts such as spreadsheets and slide decks.

A live demo shows the agent handling wedding planning end-to-end. After users enter a prompt with constraints (dress code, venue, weather, and gift considerations), the agent takes a few seconds to set up its environment, then begins a multi-step workflow that includes asking for clarification when needed, using a text browser to scan information efficiently, and switching to a visual browser to interact with web interfaces—clicking, filling forms, and verifying that suit choices look right. The workflow continues over time, with the interface showing the agent’s computer screen and an overlay of its internal reasoning as it decides what to do next.

Under the hood, the system is built around a unified tool box: a text browser for fast reading and searching across many pages, a visual browser for interacting with page UI elements, and a terminal for running code and generating/analyzing files. Through the terminal, the agent can call public APIs and—when explicitly connected—APIs tied to private data sources such as Google Drive, Google Calendar, GitHub, and SharePoint. It can also use an image generation API to create visuals for deliverables like slide decks.

Tool selection is trained using reinforcement learning on “hard tasks” designed to force the model to use all capabilities. Early in training, the model may try to use every tool even for simple problems, but rewards for correct and efficient completion drive smarter switching. Examples given include a restaurant booking flow that starts with text browsing, then moves to visual browsing for photos and availability checks, and a creative artifact workflow that searches online first, then uses the terminal for code-based compilation and verification.

The rollout is framed as a merger of earlier OpenAI agentic products: operator (for online actions like reservations and emails) and deep research (for in-depth reporting). The rationale is complementarity—operator struggles with long article reading while deep research is weaker at interactive, highly visual web pages and authenticated access—so the agent combines strengths and adds a longer-running, universal-task orientation.

Capability claims are backed by benchmark results shared in a meta-demo where the agent generates PowerPoint slides from Google Drive data. Reported gains include near-doubling performance on Humanity’s Last Exam when tool use is enabled (42%), state-of-the-art performance on Front TMS (27%) with full tool access, strong results on WebArena (improved over a prior O3-based model), and high pass rates on BrowseComp (69%). Spreadsheet editing performance is also described as improving when the agent receives raw Excel files via the terminal.

Security is treated as a central caveat. The system faces prompt injection risks—malicious websites trying to trick an agent into entering sensitive information. Mitigations include training the model to ignore suspicious instructions, layered monitoring that can stop trajectories, and the ability to update defenses as new attacks emerge. Users are encouraged to be cautious with sensitive data and to use takeover modes for direct input when appropriate.

The product is described as launching for Pro Plus and team users immediately, with enterprise and edu targeted by month’s end, alongside an expectation that safety controls and capabilities will be refined as adoption grows.