OpenAI Launches ChatGPT Agent: Thinks and Acts Autonomously, Picking Tools from a Skill Library 71

Openai has launched the ChatGPT agent, which can think and act autonomously, select tools from a skill library, and complete tasks with the help of the web. It combines Operator's ability to interact with websites, the advantages of deep research in integrating web information, and the conversational capabilities of ChatGPT to form a unified system that can turn users' ideas and questions into concrete actions, such as conducting financial research, filling out online spreadsheets, and making slides.

Main Features and Characteristics

✅ Autonomy and Collaboration: It can independently perform complex tasks such as browsing websites, selecting dates, filtering results, prompting users to securely log in, and running code. Users always remain in control and can interrupt, take over the browser, or stop tasks before important operations are executed.

✅ Multi-tool Integration: It combines the strengths of Operator and deep research, with access to a visual browser, a text-based browser, and direct API access, enabling it to select the optimal path to complete tasks.

✅ Dynamic Learning and Optimization: It learns dynamically while performing tasks, adjusting its working approach by optimizing speed, accuracy, and efficiency.

✅ Iterative, Collaborative Workflow: Users can interrupt the conversation at any time to clarify instructions, adjust task direction, or guide results. The agent proactively asks for more details when necessary to ensure tasks align with user goals.

✅ Safety and Privacy Protection: Before performing sensitive or important operations, it explicitly obtains user authorization; it is trained to resist malicious instruction manipulation; users have full control over data handling, input data is processed securely and not stored on servers.

Performance and Evaluation

✅ Scored 43.1 points on "Humanity's Last Exam," a new high.

✅ On the DSBench test, it surpassed previous state-of-the-art models, with data analysis performance outperforming humans.

✅ On the SpreadsheetBench platform, it set a new industry-leading standard, with performance more than doubling that of GPT-4o.

✅ In internal benchmarks, its ability to handle investment banking analyst tasks significantly outperformed deep research and the o3 model.

✅ On the WebArena benchmark, it performed better than the CUA powered by o3.

✅ On BrowseComp, it set a new SOTA record with 68.9 points, xx percentage points higher than deep research.

Practical Application Scenarios

✅ Work Scenarios: Automatically handle repetitive tasks, such as converting screenshots into editable presentations, rescheduling meetings, planning and booking outings, and updating spreadsheets with new financial data (while maintaining original formatting), etc.

✅ Personal Life Scenarios: Planning and booking travel, designing and booking dinners, finding professionals and arranging appointments, etc.

How to Use

Through the editor tool's dropdown menu, select "Agent Mode" at any stage of the conversation to activate. Describe the task you want to complete. During execution, the on-screen voice narration will display the operation flow in real time, and users can interrupt and take over the browser at any time. You can also set completed tasks to repeat automatically, such as automatically generating weekly metric reports every Monday morning.

Access Permissions

• Starting today, it is available to Pro, Plus, and Team users; Enterprise and Education users will have access in July.

• Pro users can perform nearly unlimited tasks per month, other paid users get 50 tasks per month, additional usage can be obtained through credits.

• It can access users' connectors; after authentication, read-only connectors allow viewing information and performing actions, operations on websites require logging in through browser takeover.

Key Questions

✅ What are the core capabilities and advantages of the ChatGPT agent?

Its core capability is the ability to think and act autonomously, select tools from the skill library, leverage the web and its own virtual computer to handle tasks, smoothly switch between reasoning and execution, and independently complete complex tasks. The advantage lies in integrating the strengths of Operator, deep research, and ChatGPT into a unified system that can turn ideas into actions, while users always remain in control—permission is obtained before important operations, and users can interrupt, take over, or stop tasks at any time.

✅ How does the ChatGPT agent perform in different evaluations?

It scored a new high of 43.1 on "Humanity's Last Exam"; surpassed state-of-the-art models on DSBench, with data analysis outperforming humans; set industry leadership on SpreadsheetBench, achieving an overall soft limit of 45.5% when supporting .xlsx format; outperformed deep research and the o3 model in handling investment banking analyst tasks in internal tests; performed better than CUA powered by o3 on WebArena; set a SOTA record with 68.9 points on BrowseComp.

✅ How can users use the ChatGPT agent, and what are its availability and safety regulations?

For usage, activate by selecting "Agent Mode" through the editor tool's dropdown menu, describe the task, and you can also set tasks to repeat automatically, with the ability to interrupt and take over during execution. For availability, Pro, Plus, and Team users have access starting today, Enterprise and Education users in July; Pro users get nearly unlimited tasks per month, other paid users get 50 per month, with additional usage available through credits. For safety, user confirmation is required before sensitive operations, critical tasks require user supervision, high-risk tasks will be refused, it can resist adversarial attacks and misuse, protect user data privacy, and users have full control over data handling.

The ChatGPT agent launched by OpenAI is still in an early stage and can handle many complex tasks, but may make mistakes. The slide generation feature is in testing, and format and detail handling may be rough; additionally, the spreadsheet upload feature is currently not available for slides. In the future, its efficiency, depth, and versatility will continue to improve, interactions will become smoother, and the level of supervision required from users will be adjusted to balance safety and practicality.

Related link: https://openai.com/zh-Hans-CN/index/introducing-chatgpt-agent/