OpenAI Introduces a General-Purpose Agent in ChatGPT: What It Means for Users and the Industry

OpenAI has launched the ChatGPT Agent, its most advanced digital assistant to date, designed to serve as a universal helper capable of taking over and automating tasks for users.

What is ChatGPT Agent?

ChatGPT Agent is a tool that merges capabilities from previous OpenAI solutions such as Operator (which can independently search websites) and Deep Research (for researching and summarizing information from multiple sources). For users, this means you can delegate a wide variety of tasks simply by typing in natural language, including:

• Automatically managing your calendar
  
• Creating and editing presentations
  
• Running code
  
• Connecting to other apps like Gmail and GitHub to quickly retrieve relevant information

How Does It Work, and Who Can Use It?

The agent is available to subscribers of OpenAI’s Pro, Plus, and Team plans. To activate it, simply select “agent mode” in the ChatGPT menu.

Its major strength lies in integrating multiple skills:

• It can access applications via ChatGPT connectors
• It has the ability to use a terminal for code execution
• It can use APIs to interact with specific apps

Some typical tasks it can handle include planning and shopping for ingredients for a meal or analyzing competitors and creating a presentation, activities which require complex information processing, planning, and practical execution.

Performance and Safety Measures

OpenAI states that the new model significantly outperforms its predecessors:

• It scored 41.6% on Humanity’s Last Exam, roughly double previous o3 and o4-mini models
• On the FrontierMath test, it achieved 27.4% with tool access, far above previous best scores such as o4-mini’s 6.3%

Given its increased capabilities and potential risks, OpenAI has paid special attention to safety:

• A real-time monitor checks all user input
• Two-level monitoring for biology-related queries: a classifier first detects biological content, and then a secondary check ensures responses cannot be misused
• The agent’s memory is disabled to prevent the possibility of leaking sensitive data through advanced attacks

Challenges and Limitations

Although the agent sounds impressive, earlier experiences with similar digital tools show they can struggle with complex real-world tasks and may be prone to errors. OpenAI itself acknowledges that these technologies have yet to fully deliver on the ambitious visions set out by tech leaders, but believes this new release brings them closer to that goal than ever before.