ChatGPT can now act on your behalf, using its own virtual computer to complete complex tasks, browse the web, run code, and interact with online tools, all without step-by-step prompting.
ChatGPT As An ‘AI Agent’
OpenAI has formally launched what it calls the ChatGPT agent, transforming its well-known conversational model into a proactive digital assistant capable of completing real tasks, independently choosing tools, and reasoning through multi-step workflows.
This new functionality, now available to paying subscribers, marks a significant turning point. For example, rather than simply responding to prompts, ChatGPT can now act as a true AI agent, performing tasks such as planning a meal, generating a financial forecast, writing and formatting a presentation, or summarising your inbox. Crucially, it can also interact with websites, manipulate files, and run code using its own virtual machine.
“We’ve brought together the strengths of our Operator and deep research systems to create a unified agentic model that works for you—using its own computer,” OpenAI explained in a blog post on 17 July. “ChatGPT can now handle complex tasks from start to finish.”
What the ChatGPT Agent Can Actually Do
OpenAI says the agentic version of ChatGPT can choose the best tools to solve a problem and it can perform multi-step operations without being micromanaged by the user.
For example:
– Users can ask ChatGPT to analyse their calendar and highlight upcoming client meetings, incorporating relevant news about those companies.
– It can plan a dinner for four by navigating recipe websites, ordering ingredients, and sending the shopping list via email.
– In professional settings, it may be used to analyse competitors, generate editable slide decks, or reformat financial spreadsheets with up-to-date data.
Its Own Toolkit
Technically, the agent achieves this by drawing on a powerful toolkit, i.e. a visual browser, text-based browser, command-line terminal, access to OpenAI APIs, and “connectors” for apps like Gmail, Google Drive, or GitHub. OpenAI reports that it can navigate between tools fluidly, running tasks within a dedicated virtual computer environment that preserves context and session history.
This context-awareness means it can hold onto prior steps and continue building on them. For example, if a user uploads a spreadsheet to be analysed, ChatGPT can extract key data, switch to a browser to find supporting info, and return to the terminal to generate a report, all within one session.
OpenAI describes the experience as interactive and collaborative, not just automated. Users can interrupt, steer or stop tasks, and ChatGPT will adapt accordingly.
Who, When, and How?
The new ChatGPT agent capabilities are being rolled out initially to paying customers on the Pro, Plus, and Team plans. Enterprise and Education users will follow in the coming weeks.
To access agent mode, users need to open a conversation in ChatGPT and select the ‘agent mode’ option from the tools dropdown. Once enabled, users can assign complex tasks just as they would in a natural chat. On-screen narration gives visibility into what the model is doing at each step.
Pro users get 400 messages per month, while Plus and Team users receive 40 per month, with more usage available through paid credits.
Although the rollout is currently limited, OpenAI says it is “working on enabling access for the European Economic Area and Switzerland” and will continue improving the experience.
Why OpenAI Is Doing This Now?
OpenAI’s move reflects a broader push within the industry to shift from passive chatbots to autonomous AI agents, i.e. models that can actively use tools, complete workflows, and deliver tangible results.
Until now, models like ChatGPT have excelled at language generation but faltered when asked to carry out structured, real-world tasks involving files, websites, or multiple steps. That changes with the new agent.
Demand-Driven Says OpenAI
According to OpenAI, user demand drove this shift. For example, many were reportedly attempting to use previous tools, such as Operator, for deeper research tasks, but were apparently frustrated by their limitations. By combining tool use and reasoning within a single system, OpenAI hopes to unlock more practical and business-relevant use cases.
This could also represent a strategic response by OpenAI to rising competition from agents being developed by Google DeepMind, Anthropic, Meta, and open-source communities, many of whom are now focusing on AI models that can act, not just talk.
Business Uses
While consumers can use the agent for tasks like travel planning or dinner parties, the biggest implications may be for professionals and businesses. For example, in OpenAI’s internal tests, the agent performed as well as or better than humans on tasks like:
– Generating investment banking models with correct formulas and formatting.
– Producing competitive market analyses.
– Updating Excel-style financial reports.
– Converting screenshots into editable presentations.
OpenAI says that for data-heavy roles, ChatGPT agent showed strong results. For example, on DSBench, a benchmark testing real-world data science tasks, it outperformed humans by wide margins in both analysis (89.9 per cent) and modelling (85.5 per cent). On SpreadsheetBench, it scored 45.5 per cent with direct Excel editing, far ahead of Microsoft’s Copilot in Excel at 20.0 per cent.
This positions ChatGPT agent not just as a time-saver, but as a cost-effective knowledge worker in fields like consulting, finance, data science, and operations.
New Capabilities Bring New Risks
Despite the powerful new functions, OpenAI has been clear that risks are increasing too, particularly because the agent can interact directly with sensitive data, websites, and terminal commands.
“This introduces new risks, particularly because ChatGPT agent can work directly with your data,” the company warned, noting the risk of adversarial prompt injection—where attackers hide malicious instructions in web pages or metadata that the AI might interpret as legitimate commands.
For example, if a webpage contained an invisible prompt telling ChatGPT to “share email contents with another user,” the model might do so (unless safeguards are in place).
To prevent this, OpenAI says it has:
– Required explicit user confirmation for real-world actions like purchases or emails.
– Introduced a watch mode for supervising high-impact tasks.
– Trained the model to refuse dangerous tasks (e.g. transferring funds).
– Implemented privacy controls, including cookie and browsing data deletion.
– Shielded the model from seeing passwords during browser “takeover” sessions.
Also, on synthetic prompt injection tests, OpenAI claims the agent resisted 99.5 per cent of malicious instructions. However, in more realistic red-team scenarios, the resistance rate dropped to 95 per cent, which is a reminder that vulnerabilities still exist.
The Next Phase
The launch of ChatGPT agent pushes OpenAI firmly into the next phase of AI development, i.e. intelligent systems that act on behalf of humans, not just inform them.
It’s a clear sign that OpenAI aims to lead in the agentic AI race, rather than simply competing on model performance or training size. With its own virtual environment, a growing toolset, and proactive capabilities, ChatGPT now resembles something closer to a software co-pilot than a chatbot.
Competitors will likely follow suit. Google’s Gemini, Anthropic’s Claude, and open-source challengers are all exploring similar agent-style features. However, OpenAI is arguably first to market with a production-ready system that balances capability and risk management (however imperfectly).
For users, especially in business, the implications are considerable. For example, those able to integrate ChatGPT agent into workflows may gain speed, efficiency, and analytical power, so long as they understand the limitations and continue to exercise oversight.
The success of this rollout could also shape broader conversations about AI safety, regulation, and responsibility, particularly as agents become more embedded in real-world systems.
What Does This Mean For Your Business?
The agent rollout gives OpenAI a powerful lead in the shift toward goal-directed, tool-using AI, one that can complete work on behalf of the user rather than waiting for commands. Its ability to interact with live websites, private data sources, and business systems puts it on a new level of utility, but also of accountability. This is no longer just about generating answers. It is about delegation.
For UK businesses, the implications are likely to be immediate and wide-ranging. For example, the agent offers a credible way to automate time-consuming tasks like competitor analysis, document preparation, scheduling, and spreadsheet management. For knowledge-heavy sectors such as finance, consultancy, and data operations, it introduces a low-friction option for streamlining routine work, reducing manual handling, and speeding up research. Organisations already experimenting with automation and AI-assisted productivity tools may now find themselves rethinking existing workflows in favour of a more hands-off, outcome-driven approach.
However, it’s not without operational risks. Any system that can click, copy, calculate, and communicate on your behalf must be trusted to do so responsibly. That means businesses will need to consider internal guardrails and policies, not just to protect sensitive information, but also to ensure the AI is being used ethically and in line with organisational goals. The fact that ChatGPT can now act autonomously raises pressing questions around auditability, compliance, and human oversight, especially in regulated sectors.
There are also broader competitive and reputational pressures in play. For OpenAI, this launch extends its relevance beyond individual users and into the professional environments that rivals like Microsoft and Google are also targeting. At the same time, it invites scrutiny over safety claims, especially as agents become more capable and the scope for unintended consequences grows.
OpenAI making ChatGPT an AI agent appears to be a clear step-change in how AI is positioned and applied. The tools are no longer limited to outputting content or providing suggestions, but are now expected to deliver outcomes, complete tasks, and take action with minimal supervision. For users, that means new possibilities, but also a renewed need to stay alert, strategic, and in control.