OpenAI Launches ChatGPT Agent, Entering the AI Agent Arena

OpenAI Enters the AI Agent Race with ChatGPT Agent

OpenAI has finally joined the heated competition in the field of AI Agents with the launch of ChatGPT Agent. This new product enables the popular ChatGPT chatbot to represent users in performing multi-step tasks, opening up vast possibilities for automating business processes and executing complex tasks with greater efficiency.

ChatGPT Agent allows users to delegate tasks such as scheduling meetings, creating comprehensive presentations, and even using specialized tools like code terminals or web browsers to complete their tasks. This ability to interact with diverse tools gives the Agent great flexibility in handling various types of required tasks.

Current Limitations and Accessibility

Currently, this new feature is only available to users who pay for OpenAI's services. Pro subscribers can use up to 400 smart prompts per month, while Plus and Team subscribers are limited to 40 times per month. Access to this feature is expected to be provided to enterprise and educational institution users in the coming weeks.

Timing of Launch and Competition

The launch of ChatGPT Agent comes after significant investments from major technology companies like Google, Salesforce, and HubSpot in the field of AI Agents over the past 12 months. It also coincides with OpenAI's recent announcement of launching its own browser, in a move similar to Perplexity AI. This timing demonstrates the intense competition in this field and OpenAI's eagerness to stay at the forefront.

How Does ChatGPT Agent Work?

According to OpenAI, ChatGPT Agent combines three key benefits: the Operator's ability to interact with websites, the ability to integrate in-depth information, and ChatGPT's intelligence and conversational fluency. Users can give ChatGPT instructions such as "Check my schedule and provide me with a brief summary of upcoming client meetings based on the latest news" or "Analyze three of the main competitors and prepare a slide presentation".

The Agent operates on its own virtual computer and can navigate between multiple tools, including a text browser, a visualization browser, a code terminal, and third-party application connectors such as Gmail and GitHub. OpenAI claims that this combination enables the model to "move seamlessly from inference to action, fully processing complex workflows".

ChatGPT Agent is specifically designed for collaboration. Users can interrupt the Agent, take over the browser, or reassign tasks as they progress, and ChatGPT will adjust accordingly. The company wrote in its blog: "It will resume where it left off and be updated with new information, but it will not lose its previous progress".

Current Limitations and Security Concerns

However, the tool does not currently support real-time responses. In the demonstration, ChatGPT Agent took about 25 minutes to create a slide presentation based on the company's financial reports. Product Manager Yash Kumar said that many tasks can be "started in the background and then come back to check them later". Research Director Isa Fulford added:

"Even if it takes 15 minutes or half an hour, this is a significant improvement in efficiency compared to the user doing it manually".

For safety reasons, some features (including memory) were disabled in the initial release. Kumar told The Verge: "It's not because we think it's unsafe, but we're taking extra precautions".

ChatGPT Agent represents a new benchmark for AI, but it also raises safety concerns. OpenAI stated that the model that supports ChatGPT Agent significantly outperformed previous technical versions in many benchmark tests. But this release is accompanied by higher risks. OpenAI classified the model in its internal "preparedness framework" as having "high biological and chemical capability", and this classification applies to cases where "models may exacerbate existing pathways to serious harm". OpenAI said: "Although we have no conclusive evidence that the model can significantly help non-specialists cause serious biological harm - which is the standard we have set for 'high capability' - we are being cautious and immediately implementing the necessary protective measures".

The company said it has taken new measures to reduce the risk of misuse in the real world, including requiring users to explicitly confirm before performing irreversible actions, and a "monitoring mode" that requires users to oversee high-risk tasks (such as sending emails or interacting with financial websites). OpenAI has also introduced safeguards to prevent prompt injection. Prompt injection is where malicious instructions are hidden in content that the Agent may encounter online. The company wrote:

"Because ChatGPT Agent can take direct action, a successful attack could have a greater impact and carry higher risks".

However, OpenAI emphasizes that today's release is "just the beginning". The company plans to continue improving and refining the Agent's capabilities over time. OpenAI said: "We expect the efficiency, depth and diversity of ChatGPT Agent to continue to improve, including achieving smoother interactions".