From Conversation to Execution, the Beginning of ''Complete Automation'' Handled by a Single Agent
On July 17, 2025, OpenAI announced ChatGPT Agent — an AI that browses the web, executes code, creates presentations, orders ingredients, and plans meals autonomously. ChatGPT Agent operates within a unified virtual computer environment, capable of autonomously handling complex real-world tasks from start to finish. Five core capabilities: (1) Visual and text-based browser use — clicks websites, filters data, fills forms, handles login procedures like an actual user, automating online shopping, article collection, and reservations; (2) Terminal and API integration — code execution, file analysis, connections to Gmail, Google Calendar, GitHub for complex multi-step tasks; (3) User control and safety — mandatory user permission for sensitive actions (login, payment, personal information access); user can take browser control or pause tasks at any time; (4) Multi-step scenario execution — handles complex chains like "summarize my meeting schedule, find related news, and create a slide deck" by sequentially accessing calendar, researching news, and building presentation; (5) Human approval for irreversible actions — purchases, email sends, and other permanent actions always require explicit confirmation before execution. Availability: ChatGPT Plus subscribers in the US, with broader rollout planned. Industry significance: ChatGPT Agent represents the shift from AI as information tool to AI as action-taking partner — directly competing with emerging "agentic AI" products from Anthropic (Claude's computer use), Google (Project Mariner), and Microsoft (Copilot Actions), marking the beginning of the "agentic AI" commercial era.
