OpenAI has launched ChatGPT Agent, a new feature that enables the AI to proactively complete multi-step tasks on behalf of users using its own virtual computer. This blends the best of OpenAI’s earlier tools, Operator and Deep Research, into a unified system that interacts with the web, writes code, manipulates files, and creates deliverables, such as slideshows and spreadsheets.

This launch expands ChatGPT’s role from a conversational assistant to a semi-autonomous agent. However, users retain control with the ability to interrupt, redirect, or oversee any action.

With built-in browser access, terminal use, API connectivity, and app integration, ChatGPT Agent can navigate between reasoning and action. Its capabilities will initially be available to Pro, Plus, and Team users, with plans to expand further.

Why It Matters: This is a new chapter for AI utility, moving beyond static chatbot interactions into live, actionable assistance. If successful, this could change how professionals and individuals manage digital tasks, making AI act more like a reliable partner. However, it also introduces new safety and privacy risks, prompting OpenAI to implement extensive safeguards.

• From Static Models to Active Agents: Earlier AI models were mostly reactive and would generate responses or content based on prompted questions. ChatGPT Agent represents a major evolution by taking initiative in navigating the web and interacting with tools to complete multi-step tasks from beginning to end.

• A Rich Set of Tools: Users can activate “agent mode” to let ChatGPT operate through a browser, terminal, or API connection. It can access apps like Gmail or GitHub through connectors to manipulate files, and shift between tools as needed, while keeping the task context intact. The agent can even notify you when it finishes a job, making it more useful for ongoing work.

• Top Performance Across Benchmarks: The agent sets new records on several evaluations, including scoring 41.6% on Humanity’s Last Exam and 27.4% on the complex FrontierMath benchmark. It also significantly outperforms previous models on tasks like data science and even investment banking modeling, indicating its capabilities to complete general and specialized knowledge work.

• Smarter with Supervision: While the agent can work more independently than past models, it’s designed for collaboration and oversight. It pauses for user input before taking significant actions and adapts based on feedback mid-task. This is a step towards a more supervised AI that acts more like a team member.

• Safety Scaling Alongside Capability: The risks evolve as the technology does. OpenAI has adapted its safety strategies accordingly by introducing features like live monitoring and disabling memory. By increasing these precautions, the company is acknowledging that guardrails must be expanded as AI becomes more powerful.

Go Deeper —> Introducing ChatGPT agent: bridging research and action – OpenAI

OpenAI launches a general purpose agent in ChatGPT – TechCrunch