ChatGPT Agent Launched: Capable of Performing Dozens of Tasks Independently

ChatGPT Agent, capable of performing dozens of tasks independently, from scheduling appointments to running code, writing research reports, and creating presentations, has been rolled out.

OpenAI has introduced ChatGPT Agent (also usable as an “Agent” or “Broker”) feature, transforming AI from a passive chatbot into an active digital assistant. This new tool can perform multi-step operations on the user’s computer on its own. ChatGPT Agent is capable of executing complex tasks such as Browse websites, gathering information, creating presentations, and running code.

OpenAI has begun making this tool available to ChatGPT Pro, Plus, and Team users. The tool can be activated through “agent mode” in the ChatGPT interface or directly with the “/agent” command. Broader access for enterprise and educational plans is planned towards the end of summer. Pro users have virtually unlimited task rights per month, while other paid users receive 50 tasks per month and have additional usage rights with flexible credit options.

You Write, It Does the Rest

ChatGPT Agent performs complex tasks based solely on natural language commands written by users. For example, if a user says, “Plan a Japanese breakfast for four people and buy the ingredients,” the agent first researches suitable recipes, then creates the ingredient list, and becomes capable of ordering them from online markets. Similarly, requests such as “Analyze three competitors and prepare a presentation” can be transformed into research reports supported by slides. Meanwhile, ChatGPT Agent is not entirely autonomous. It will ask for your permission when an irreversible decision needs to be made (e.g., placing an order, sending an email). To perform these tasks, the capabilities of OpenAI’s previous “agents,” “Operator” and “Deep Research” systems, have been combined under one roof. Operator was an interface controller that could click buttons and fill out forms. Deep Research acted as a research engine that could summarize dozens of different web sources. ChatGPT Agent, by combining these two systems, has gained a much wider range of tasks.

Gmail and GitHub Integration, Terminal Access, and API Usage

The new agent can connect to external services like Gmail and GitHub through ChatGPT Connectors. This allows users to access and analyze information in their emails or collect data from code repositories on GitHub. At the same time, the system has terminal access; meaning it can write, run, and interpret the output of code.

Additionally, thanks to API integration, the agent can communicate with different web services, undertaking more tasks. Thus, it can not only generate content but also perform interactive actions.

At the Peak of Its Performance

According to OpenAI, ChatGPT Agent operates with one of the most powerful models the company has developed to date. The model achieved a 41.6% success rate in the challenging “Humanity’s Last Exam.” This rate is almost twice as good compared to OpenAI’s o3 and o4-mini models. It also showed a 27.4% success rate in the FrontierMath benchmark test. This rate is approximately four times the previous record.

Security and Limitations

OpenAI has significantly increased security measures due to the new system’s advanced capabilities. ChatGPT Agent has been classified as “highly capable” in biological and chemical fields and is equipped with special security measures accordingly.

Every command from the user is filtered in real-time. If the prompt contains a topic related to biology, a second control mechanism is activated. This system evaluates whether the response is potentially harmful.

Furthermore, the memory feature in ChatGPT Agent has been disabled. Normally, ChatGPT can remember previous conversations and provide contextual answers. However, this feature is kept off in agent mode as it could lead to the leakage of sensitive data.

In the final part of the content, it’s also necessary to mention some security concerns. As we said, ChatGPT Agent can access most areas. For example, if you give this agent the command:

“Check my emails received overnight and do everything necessary to address them, do not ask follow-up questions.”

The agent will read your emails and do what’s necessary for you based on the command, but this also poses a very significant risk. What if one of the emails carries a malicious purpose? An email from an untrusted source could result in the AI leaking sensitive information itself. Therefore, when using ChatGPT Agent or similar tools, I recommend giving these tools as little access permission as possible.

You Might Also Like;