Want a smart insight into your inbox? Sign up for our weekly newsletters to get the only thing that is important to enterprise AI, data, and security leaders. Subscribe now
Is not letting up Delay in its open source AI model Slow it when sending other features.
Today, The company is exposing Chat GPT AgentA feature that allows your AI chatboat to browse the web independently, do extensive research, use new virtual computer to create new files for its human users.
Then come? Chatgupat now gets her Own PC? And can it use this PC to download your own, human user, accounts and download or send goods for you?
According to the open, this is true, at least in virtual sense. As the company explains:
AI Impact Series returning to San Francisco – August 5
The next step of the AI is here – are you ready? Block, GSK, and SAP leaders include for a special look on how autonomous agents are changing enterprise workflows-from real time decision-making to end to automation.
Now secure your place – space is limited:
“The model can choose to open a page using a text browser or visual browser, can download a file from the web, run the command in the terminal, and then see the output in the visual browser.
How to use Chat GPT Agent
Users can add the agent by clicking the ‘Tools’ button in the Chat GPT prompt entry box, opening the dropdown menu, and selecting ‘Agent Mode’ from the available options.
Then, when it is turned on, explain a task in a simple language, and the agent can take it into the web and local app environment, connecting the reasoning with the actions that only a human user can perform on his machine manually.
The Chat GPT agent can connect your personal or business Gmail and Gut Hub apps, so it can pull useful information from your accounts – emails or code to help you do the tasks you do. This information can be connected to the third -party application programming interface (API) to pull and use the applications and services connected through them.
If a website you need to log in, you can safely do it through a particular browser view, which allows the agent to dig deep and handle more personal tasks, such as checking your inbox or filling out your form.
Where the operator couldn’t – offline
The new Chat GPT agent, released in January 2025, produces and spreads from the “Operator” agent Openi, which allows Chat GPT to browse the web and fill out forms, order, order, and other web -based tasks in a private “headless browser”, which is based on the web -based web.
However, the operator was only limited to communicating with websites and web-based applications-not a program that can run locally on the PC, such as the software to create a spreadset tablets and slides-deck presentation.
The new chat GPT agent can browse websites, interact with online forms, run the code, analyze data, and finished outpts – such as editing offers or spreadsheets – is fully based on user instructions.
This mask comes on the heels of a report that has been published by the Independent Subsection Tech Industry website Information It suggested that the Openi will upgrade the Chattgopt Microsoft’s Office Software Applications (such as Excel, Word, PowerPoint, etc.) to become more directly competitors
Operator and integrate deep research into an agent
In fact, the Openi has placed the Chat GPT agent as his two ex-agents-operators and deep research integration, the latter was introduced in February 2025, which finds the information to find and compile information in long and depth reports. As Open AI writes in a blog post:
“The operator could not have a deep dive in the analysis or wrote detailed reports, and deep research websites could not interact with the websites to improve the results or access the content needed to confirm the user’s verification.
The previous operator tool will be eliminated phasedly, but users can still access deep research through dropdowns in the Chat GPT interface.
Whether use a visual browser to communicate with a website or terminal to run a coded code, the agent runs without interruption between the tools within the same session.
It supports rivals and reports to planning, summarizing emails, summarizing emails, or baking of appointments to a wide range of issues of use.
Consumers can interfere with, redirect or withhold a work at any time, with the agent picking up where they leave.
Availability and access
Starting today, Chat GPT’s $ 200 per month “Pro” level users will have full access to Chat GPT agent with a monthly quota of 400 messages.
Chat GPT Plus ($ 20 per month) and the team ($ 30 every month) will access with 40 messages a month, in the next few days. Additional use is available through credit -based options.
Openi said in a joint release with Venture Bat under Embigu that its chat GPT Enterprise and Education users will access this feature in the coming weeks.
For now, this feature is not yet available in Europe or Switzerland, there is no doubt that the residents there are disappointing.
Made with the forehead safety and control
Given that the agent can now take action on behalf of users-including the login websites or with associated apps-Openai has introduced extensive safety measures.
These include technical safety measures to limit user verification, active surveillance for sensitive tasks, and limit unnecessary behavior before taking action.
Key reservations include:
- Verification indicated Before steps such as submitting a form or sending email
- Watch modeWhen the user becomes disabled, stops implementation
- Deny of high -risk tasksSuch as a violation of financial transfer or privacy
- Not the maintenance of memory During the agent session
Rating high -risk domain rating
According to its preparation framework, Openi Chat is behaving as a GPT agent Superiority System in biological and chemical domains.
Although there is no direct evidence of abuse, the company is activating its strong security protective measures due to caution.
These include better refusal training, red -teaming by biochemical experts, and better detection systems.
Remember that the recent information of Anthropic has released about his new Claude Ops 4, and Other surveys of advanced AI modelsIt shows that when access to external tools and applications such as email, they can take these steps morally and morally in some cases, but it may compromise with the user, such as emailing government agencies or journalists to do suspicious wrongdoing.
The model believes that he is acting like a “whistleblower” but in fact, the user’s privacy, security, and the proprietary information and the authorities can inform the wrongdoing where no one is, or where it is suspicious.
Strong performance on real -world tasks
Chat GPT agent is not much in theory. It set up a new high score in the final examination of humanity with 44.4 using parallel rollout methods, and achieved 27.4 % on a hard frontiermate benchmark.
On the spreadsheet bench, he scored 45.5 % of the Co -Pilot in Excel’s performance.
Current limits and next steps
Some features, such as slideshow generation, are still in betting and can feel fundamental in formatting or are slightly different between previews and exported files in the app.
Openi is actively training the next repetition of this feature to improve Polish and layout.
The launch of the Chat GPT agent indicates how consumers interact with AI – from asking questions to assign the full work.
With its reasoning, the ability to act and supply, the open is bet that consumers are quickly want AI to help them not only help them, but also work for them. Although the company emphasizes that the agent is still being manufactured, this launch sees the more interactive, an action -based future for AI.
