ChatGPT’s New “Agent”: A Promising but Imperfect Assistant

Imagine assigning a task and having an agent powered by ChatGPT “take care of the work for you.” This is the ambitious promise of OpenAI’s new tool, launched recently and available only to paying subscribers. The agent is currently unavailable in the European Union.

According to OpenAI, this agent can “look at my calendar and then brief me on my upcoming client appointments based on the news” or “automatically rearrange my appointments.” This requires granting ChatGPT access to some of the digital services you use – like your email or calendar – by creating what the company calls a “connector.”

OpenAI claims the agent can also “fill out online forms or create polished presentations.” To do this, it can gather information from the internet, potentially requiring user authentication on subscribed websites. The agent will then navigate the site independently, clicking on buttons it deems appropriate. A task can take up to several minutes, with a notification appearing on the user’s smartphone upon completion.

Intense Competition, Questionable Efficiency

OpenAI is not alone in pursuing agents capable of performing complex tasks. Many other major American tech companies have been working towards this goal. The first consumer tools emerged recently, with Anthropic announcing its Computer Use system, followed by Microsoft’s Copilot Actions, and Google’s Gemini in Chrome, which faced criticism.

Developing this type of tool appears challenging. Apple, for example, touted the imminent arrival of an AI-powered “agent” but later remained silent about it during a presentation of future AI features. The release date of Apple’s agent is currently unknown.

OpenAI acknowledges that its ChatGPT agent is not yet fully developed: “it can make mistakes,” the company states. However, they assert that its performance is significantly better than the experimental Operator agent launched previously, which was also criticized. To improve it, the company merged it with Deep Research, its in-depth research tool.

Security Concerns

To address customer concerns, OpenAI emphasizes the security measures implemented in its new agent. However, it acknowledges that its use increases the risks associated with ChatGPT, particularly due to the sensitive personal information it can access. The company recommends careful consideration before granting access to calendars or email inboxes.

OpenAI also warns that “since the ChatGPT agent can perform direct actions, successful cyberattacks can have a greater impact.” Therefore, it has trained its agent “to refuse high-risk tasks like bank transfers.” The company demonstrates an automated task: the ChatGPT agent creates a sticker for a business, finds an online printer, and orders several hundred copies. However, the robot stops before finalizing the purchase, handing over the task to the human user, who is the only one authorized to enter bank details.

---

Enjoyed this post by Thibault Helle? Subscribe for more insights and updates straight from the source.