OpenAI Introducing ChatGPT agent: bridging research and action

The ChatGPT agent, introduced by OpenAI in July 2025, is a new unified agentic system that enables ChatGPT to think and act autonomously by proactively choosing from a toolbox of agentic skills to execute complex, multi-step tasks on your behalf using its own virtual computer.

Core Capabilities:

• Autonomous task execution: ChatGPT can navigate websites, interact with web pages (click, scroll, type), log in securely when needed, run code, conduct complex analysis, and produce editable outputs such as slideshows and spreadsheets.
• Unified system integrating previous tools: It combines the web interaction strength of Operator, deep synthesis skills of deep research, and ChatGPT’s intelligence, offering seamless transitions within a single conversation from casual inquiry to detailed task automation.
• Multitool environment: Equipped with multiple tools including:
  • Visual browser for graphical browsing,
  • Text-based browser for data-heavy queries,
  • A terminal for code execution,
  • Direct API access,
  • Connectors for apps like Gmail and GitHub to access contextual user data securely.

User Control & Safety:

• Users retain full control over the agent:
• • ChatGPT requests permission before performing any consequential action.
  • Users may interrupt, take over the browser, pause, or stop tasks at any time.
• Strong risk mitigation against prompt injection and other adversarial attacks has been implemented.
• Privacy controls allow users to delete browsing data and log out of sessions; credentials and sensitive data entered during browser takeover sessions are never stored by the model.

Practical Applications:

• Automates everyday and professional workflows such as:
• • Calendar briefing based on news,
  • Planning and purchasing groceries,
  • Competitor analysis with slide deck creation,
  • Automating financial modeling,
  • Converting screenshots to presentations,
  • Booking travel and appointments,
  • Editing complex spreadsheets, where it significantly outperforms other models.

Performance and Benchmarks:

• Achieves state-of-the-art results across benchmarks measuring web browsing, economic knowledge work, data science, spreadsheet editing, and complex mathematical problem solving.
• Outperforms prior models and often matches or surpasses human performance in professional tasks.

Availability:

• Available to Pro, Plus, and Team users, activated via the tools dropdown in ChatGPT by selecting “agent mode” at any point during a conversation.

Safety and Ethical Considerations:

• Classified as having high biological and chemical capability risk; enhanced safeguards include threat modeling, refusal training, and expert review.
• Collaboration with biosecurity experts ensures robust safety and compliance.

In essence, ChatGPT agent represents a significant advancement toward truly autonomous AI assistants capable of complex, real-world task execution with user-controlled, transparent, and secure workflows.