Last week, OpenAI announced and launched ChatGPT Agent, giving ChatGPT both the ability to perform deep research to synthesize information and the Operator's ability to interact with websites. OpenAI showed how the new ChatGPT agent can bounce between a visual browser, a terminal, and APIs to complete the given task.
OpenAI spoke about how ChatGPT Operator, while having the ability to interact with websites, wasn't good at comprehending massive contexts and articles. Whereas, ChatGPT Deep research could comprehend large contexts and articles, but didn't have the ability to interact with websites. Hence, it was logical to combine them both into one, and thus came ChatGPT Agent.
Until now, ChatGPT could write prose or code, but it couldn't click a button. Operator handled websites, Deep Research summarized data, and the base model chatted; each function was handled and completed separately. However, the new ChatGPT agent stitches those strengths together into one super agent.
Here's a breakdown of its main features, functions, and key points:
- Unified toolbox: It features a visual browser for human-style navigation, a text browser for quick parsing, a terminal for code, and API connectors for apps like Gmail and GitHub.
- Virtual isolation: It works within its own virtual computer, preserving the context of your task even when switching between tools.
- Permission first: The agent always asks for permission before any big moves like submitting forms, spending money, or sending emails, allowing you to jump in, pause, or redirect anytime.
- Iterative workflow: You can jump in and interact with the agent mid-way as it works, clarifying instructions, changing the task, or requesting progress summaries without losing progress.
- Performance Edge: ChatGPT agent has scored high on tough tests, like 41.6% on expert-level questions or 27.4% on hard math problems, often outperforming older models.
- Real-World Applications: In a business context, it automates tasks such as updating financial sheets or booking team events. For personal use, it can plan and book your entire itinerary or schedule appointments.
This isn't just a list of new features; the performance backs it up. On multiple industry evaluations for real-world tasks, the new ChatGPT Agent has shown considerable improvements. On an internal benchmark measuring complex knowledge work, its output was comparable to or better than that of human experts in roughly half of the cases. For data science tasks, it surpassed human performance, and it scored significantly higher than tools like Microsoft's Copilot in Excel for the ability to edit spreadsheets directly.
Get Started with the New ChatGPT Agent: Step-by-Step Guide
Step 1: Getting started with the new ChatGPT Agent is easier than it sounds.
- Go to ChatGPT and make sure you have the ChatGPT Plus subscription to use this new AI agent.
Step 2: Activate the ChatGPT Agent by selecting the "agent mode" and give it a task prompt.
- For the purpose of this article, we requested a simple competitor analysis report.
Step 3: Once you hit submit, ChatGPT Agent will prepare a summary of the task it needs to perform. You can either accept it or change it according to your particular needs.
Step 4: If you are satisfied with the task summary, click 'Continue' to allow the ChatGPT Agent to set up the virtual desktop. Within that virtual desktop, the ChatGPT Agent accesses the internet for information and also interacts with websites, executing code within its integrated terminal.
Fyi: this video has been sped up!
Step 5: My task took about 21 minutes to complete, which may seem a lot, but personally, I was satisfied with the output as it had everything I wished for, even though the initial prompt I gave it was vague and not detailed.
In Conclusion:
The ChatGPT Agent is a major update combining the best of ChatGPT, deep research, and Operator to create a super agent. It is safe to say we are moving towards artificial general intelligence (AGI) as these autonomous AI agents improve, becoming more practical and powerful.
OpenAI has unified research, interaction, and execution within a controlled sandbox, making ChatGPT an all-purpose super agent. If I am right, saying this is a move towards AGI, where AI will genuinely feel useful at work for the average team, then this is the first draft.
