ChatGPT agent

Type: app · Vendor: OpenAI · Language: proprietary · License: proprietary · Status: active · Status in practice: mature · First released: 2025-07-17

Links: homepage docs

ChatGPT agent carries out multi-step tasks for a user by operating its own virtual computer with a browser, terminal, and connected tools, pausing for permission before consequential actions.

Description. ChatGPT agent is OpenAI's agentic mode inside ChatGPT that completes end-to-end tasks on a virtual computer rather than only answering questions. The agent plans steps, navigates a visual and text browser, runs code in a terminal, manipulates files, and reads from connectors such as Gmail and Google Drive. It is trained to request user permission before actions with real-world consequences such as purchases or sending email, and the user can interrupt the run or take over the browser at any point. It combines the earlier Operator browser agent, deep research web synthesis, and ChatGPT conversation in one system.

Agent loop shape. The agent runs inside an isolated virtual computer that keeps context across tools. It reasons about the goal, drives a visual or text browser and a terminal, reads files and connector data, and iterates step by step. Before any action with real-world consequences it pauses to ask the user for permission, and the user can interrupt, stop, or take over the browser at any point.

Primary use cases

autonomous web research and synthesis
multi-application task completion on a virtual computer
form filling and web navigation on the user's behalf
spreadsheet and presentation generation from gathered data

flowchart TD fw["ChatGPT agent"] fw --> p1["Computer Use (core)"] fw --> p2["Human-in-the-Loop (first-class)"] fw --> p3["Browser Agent (first-class)"] fw --> p4["Code Execution (supported)"] fw --> p5["Tool Use (supported)"]

Key concepts

Virtual computer → computer-use (docs) — An isolated environment with a browser, terminal, and file system whose state is shared across tools, so the agent can carry intermediate results from one step to the next within a single task.
Takeover mode → human-in-the-loop (docs) — A control in which the user takes over the visual browser to input sensitive information directly, leaving the agent out of the loop for that step.
Connectors → tool-use (docs) — Integrations that let the agent read from connected apps such as Gmail, Google Drive, and GitHub so it can use information from those services in a task.
Deep research (docs) — The web-research capability, carried over from OpenAI's Deep Research, that conducts extensive multi-step searches and synthesises the findings into a report or structured output.