OpenAI has introduced a universal agent in ChatGPT that can perform complex digital tasks at a user's command. The tool can create presentations, analyze competitors, and plan purchases using information from connected apps like Gmail and GitHub.
According to the company, the ChatGPT agent combines capabilities of earlier products, including Deep Research and website navigation. It interacts with users through natural language prompts and leverages a variety of tools, including a terminal and API. The new mode can be activated via a dropdown menu in the interface.
This marks OpenAI's first full-fledged step toward turning ChatGPT into an agent-based platform that can not only generate responses but also perform various actions. Previously, other technology companies attempted similar experiments, but most solutions failed to handle multi-step tasks. OpenAI claims that its new agent is functionally superior to previous developments.
The company cites use cases such as automatically planning a Japanese breakfast or conducting competitor analysis followed by creating a slide presentation. These scenarios require complex work, ranging from data collection and structuring information to managing actions in real time.
The agent is built on a model that has achieved high scores in AI model benchmarks. However, OpenAI notes that the product has potentially dangerous capabilities, including significant proficiency in handling biological topics.
In response, OpenAI has implemented several protection mechanisms. All queries are screened for biological content, and suspicious queries undergo additional verification. Additionally, the agent’s memory feature has been disabled to reduce the risk of data theft from instantaneous injection attacks. OpenAI states that re-enabling the memory feature may be considered in the future.
Despite the impressive capabilities, it remains to be seen how stable the agent is when solving problems in real-world environments. OpenAI asserts that this version is much more advanced, but its effectiveness can only be validated through large-scale user adoption.
