AI service provider OpenAI unveils the ChatGPT Agent, its latest personal assistant offering
On July 17, 2025, OpenAI unveiled a new integrated agent for ChatGPT, marking a significant evolution in AI technology. This agent is equipped to perform a variety of tasks for users, such as booking a restaurant, online shopping, checking a calendar, or manipulating files, all without leaving the chat interface.
However, the transparency and control issues surrounding this new agent are a topic of discussion. The agent's autonomous capabilities, which allow it to perform complex, multi-step tasks using a virtual computer, raise questions about user oversight. While OpenAI assures users that they remain in full control, the system explicitly requests permission before taking actions with significant consequences and allows users to interrupt, take over the browser, or stop tasks at any time.
Transparency challenges arise because the agent dynamically chooses actions from a "toolbox" and fluidly switches between reasoning and execution. This might make its internal decision process less immediately visible to users compared to traditional chat-only models. Furthermore, the integration of functionalities from previous tools like Operator and Deep Research into one unified agent can complicate the tracing of how outputs were generated.
Control issues concern how much autonomy the agent should have, especially with sensitive tasks like logging in securely or making purchases on behalf of users. Although OpenAI highlights the user's ability to intervene, this level of autonomy introduces risks if users are unaware or misunderstand when and how actions are performed on their behalf.
In summary, the complexity in understanding the agent’s decision flows due to action-reasoning integration and the combination of multiple tools into one system can obscure which tool is operating and how results are derived present transparency issues. Ensuring users can effectively supervise and interrupt agent actions at will and managing agent autonomy with sensitive tasks require strong user consent mechanisms.
Despite these challenges, OpenAI is taking steps to mitigate them. The system requires permission before consequential actions and enables real-time user oversight. The agent is also trained to refuse suspicious requests, such as bank transfers.
The ChatGPT agent is available only to paid subscribers with Pro, Plus, and Teams subscriptions. Unlike previous versions, this agent can navigate the web, interact with software, and make decisions based on user instructions. The agent can scan the user's calendar to find an open evening, then find well-rated restaurants that match the user's tastes, and present its selection in a few minutes.
This launch of the ChatGPT agent is part of a broader trend started by Google and Anthropic, which have also unveiled similar agents. However, with the rise of multifunctional autonomous agents, transparency and control issues need to be watched closely. The question of monetization for the new agent remains open. Currently, no sponsored recommendations are planned for the ChatGPT agent.
Artificial-intelligence technologies, such as the new integrated agent for ChatGPT, are raising questions about user oversight and transparency, particularly with tasks that require significant autonomy. The dynamic decision-making process and combination of multiple tools in one system can make it challenging for users to understand how results are derived and supervise the agent's actions effectively.
Moreover, the integrated agent's ability to navigate the web, interact with software, and make decisions based on user instructions suggests that controlling and monitoring sensitive tasks like secure logins or purchases will become increasingly complex, as the level of autonomy introduces potential risks for unaware or misunderstanding users.
