Openai introduces Chatgpt Agent: From Research to Real-World Automation

exist July 17, 2025Openai launches chatgpt proxytransform Chatgpt from conversation assistant to unified AI agent Ability to autonomously perform complex multi-step tasks (from web browsing to code execution) in a virtual computer environment.

Bridge previous functions

The ChatGpt proxy is built on two older tools:

• Operatorlimited web interaction is enabled – with the help of a browser’s proxy, click, scroll and form population.
• In-depth researchprovides autonomous browsing and reporting synthesis over a longer time frame.

Individually, both have limitations: operators can interface but cannot perform in-depth analysis; in-depth study can analyze but cannot interact dynamically with the site. Chatgpt proxy combines two advantages, unified browsing, tool usage, and reasoning within a single proxy architecture.

Internal architecture and workflow

The core is Virtual Computer Environment combination:

1. one Visual Browser For human websites,
2. one Text browser Optimized for structured reasoning.
3. one Shell/terminal For executing code,
4. Fusion API Connector Used for services such as Gmail or Github.

Agents keep adjusting (introduction to click buttons, run scripts or parse content), while keeping state across tools. All actions occur in a controlled proxy environment to ensure traceability and flexibility.

Sample tasks: From planning to execution

The ChatGpt proxy can handle the following tasks:

• Calendar Brief: Scan the calendar to get relevant news and summarize upcoming meetings.
• Grocery ordering: Purchase ingredients, compare prices, place orders.
• Competition Analysis: Get competitor pages, scratch data, create slideshows or spreadsheets.
• Financial Modeling: Download data, update spreadsheets, and retain formats.

These workflows involve the usage of multimodal tools: log in to the site, run scripts in the terminal, and then wrap the results into editable documents – all under your supervision.

Performance: Benchmark and human comparison

Openai reports huge benefits from multiple benchmarks:

• The final exam for humans: Pass @1 rate is 41.6% (best proxy result); parallel trials can reach up to 44.4%
• Front: Use terminal and code support, with a precision of 27.4%, performing better than the prior model.
• Spreadsheet board: The total score of XLSX edits is 45.5%, while Excel’s 20% and human scores are about 71%
• Knowledge work benchmarks provided internally: Agent tools meet or exceed expert performance approximately 50% of the time
• BrowseComp & Webarena: The new status of browsing-based tasks is 68.9%

These assessments show significant improvements in both autonomy and mission complexity.

Safety and risk mitigation

Agent autonomy introduces new risks. Openai has implemented several safeguards:

• Confirm clearly Before any consequence action is taken (e.g., purchase, release).
• Watch Mode: Certain sensitive tasks require active supervision.
• Strong Timely injection defenseincluding web prompts to detect exceptions and monitoring tool output.
• Privacy mechanism: Session-specific takeover mode without the need to retain sensitive inputs such as passwords.
• Biotherapy measures: High risk classification as bioagents, trigger enhanced threat modeling, rejection of training, real-time monitoring and vulnerability bounty systems.

These layers are designed to reduce abuse – from data leaks to task hijacking.

How to get started

Now available Pro, Plus and Team user:

• Professional users Get access immediately with 400 proxy mode messages/month.
• Add to team It will be accessed gradually in the coming days (40 messages per month).
• Business and education The tier will be carried out in the coming weeks.
• Rolling releases outside of the United States (Switzerland EEA) are underway.

You can switch to Agent Mode through the Tools menu in any conversation and describe the required workflow. The progress is narrated in real time and you can pause, take over or stop at any time.

Importance to AI’s workflow

ChatGpt proxy represents a leap from passive query response systems to active digital workers. By combining:

• Language reasoning (via GPT-4 class model),
• Process orchestration (browser, terminal),
• The execution environment of the context,

…Openai is implementing more Autonomous, reliableand Action-oriented Use cases. Although controls are critical to prevent abuse, this version expands the scope of what an AI assistant can actually do Donot just explain.

For developers and data scientists, ChatGpt proxy becomes a platform: a programmable, observable proxy capable of scratching, parsing, synthesising and exporting on demand. It provides opportunities for the next workflow in research, business automation, and personal productivity.

in conclusion

ChatGpt proxy is not only a conversational enhancement, but also a strategic hub for universal autonomous AI workflows. Its debut marks the transition from passive consultant to active agents to perform research, creation and real-world actions in a unified, controlled environment. It is expected that this will mature into the basic capabilities across AI domains.

Michal Sutter is a data science professional with a master’s degree in data science from the University of Padua. With a solid foundation in statistical analysis, machine learning, and data engineering, Michal excels in transforming complex datasets into actionable insights.