Everything you need to know about OpenAI’s new agent for ChatGPT – including how to access it and what it can do
ChatGPT agent will bridge "research and action" – but OpenAI is keen to stress it's still a work in progress
OpenAI has unveiled an AI agent for ChatGPT users that can plan your meals, research and assemble slideshows, and plan your holidays – though the company has warned the tool remains in beta and could make mistakes.
The arrival of "ChatGPT agent" follows a flurry of agentic AI launches over the last year, with Salesforce, Anthropic, and Google all capitalizing on the big industry trend.
In a blog post announcing the tool, OpenAI said ChatGPT will be able to carry out tasks autonomously on user devices, handling “complex tasks from start to finish”.
Simply put, this means users will be able to automate an array of activities, letting it step beyond research and generation of text, images, and code.
“ChatGPT now thinks and acts, proactively choosing from a toolbox of agentic skills to complete tasks for you using its own computer,” the company said.
However, OpenAI warned that the agent remains in "early stages" and is considered a beta rather than a finished product.
"It’s capable of taking on a range of complex tasks, but it can still make mistakes," the company added.
What ChatGPT agent can do
OpenAI said ChatGPT agent is powered by a "unified agentic system" that brings together three tools:
- Operator, a tool that could navigate and interact with websites.
- Deep Research, for analyzing and understanding information.
- ChatGPT, to essentially talk through it all.
"By integrating these complementary strengths in ChatGPT and introducing additional tools, we’ve unlocked entirely new capabilities within one model," OpenAI said. "It can now actively engage websites — clicking, filtering, and gathering more precise, efficient results."
OpenAI gave a few examples of the agent's capabilities, explaining users could ask it to check calendars and make briefing notes for upcoming meetings by pulling in recent news.
The tool can also analyze competitors and make a slide deck, convert screenshots or dashboards into presentations, book or rearrange meetings, and update spreadsheets with financial data.
It can even plan and buy ingredients to make a "Japanese breakfast for four” and recurring tasks can be scheduled to run automatically.
Still some teething issues
OpenAI noted that the results remain imperfect. For slideshows, exporting to Powerpoint may lead to "discrepancies"..
"At the moment, outputs can sometimes feel rudimentary in its formatting and polish, particularly when starting without an existing document," the company admitted.
Beyond business use cases, OpenAI said ChatGPT agent could be used in your personal life to plan and book travel itineraries, design and arrange dinner parties, and schedule appointments.
It can also shop for you – OpenAI suggested trying to find a Japanese-inspired vintage-style samsara lamp on Etsy priced under £200 with free shipping — and book restaurants.
How ChatGPT agent works
The system uses its own browsers, a visual version as well as text based, to complete such actions, interacting with third-party tools such as Gmail and GitHub using ChatGPT connectors.
"ChatGPT will intelligently navigate websites, filter results, prompt you to log in securely when needed, run code, conduct analysis, and even deliver editable slideshows and spreadsheets that summarize its findings," the blog post noted.
ChatGPT agent also narrates its activities so it's clear what it's up to at any given point. OpenAI added that the agent can be halted at any time and even interrupted to "clarify" instructions.
Similarly, the agent may ask for more details itself.
"ChatGPT requests permission before taking actions of consequence, and you can easily interrupt, take over the browser, or stop tasks at any point," it said.
The company added that critical tasks such as sending an email require "active supervision" and any action that's deemed high risk, such as a bank transfer, will be refused.
OpenAI said it would continue to add features and improvements to the agent tool, saying "today's launch is just the beginning."
How to get your hands on OpenAI’s new agent tool
To use the new agent tool, users with a Pro, Plus, or Team subscription can enable "agent mode" in ChatGPT's "composer" bar.
Access will extend to Enterprise and Education users in the next few weeks, the company revealed.
"We are still working on enabling access for the European Economic Area and Switzerland," OpenAI noted, without explaining the delay, though it is immediately available in the UK.
Because of the launch of ChatGPT agent, the Operator preview will be shut down in a few weeks, though the deep research function will remain accessible in ChatGPT.
Real action means real risk
OpenAI warned that giving AI the ability to take actions and access your data comes with inherent risks – a recurring talking point since the agentic AI trend burst onto the scene late last year.
Research from Gartner, for example, shows concerns about agentic AI security risks have been grumbling away among enterprise IT leaders, prompting the rise of ‘guardian agents’ to monitor those working on the frontline for businesses.
OpenAI said it’s made efforts to implement “robust controls” and safeguards when handling sensitive information on the web, for example.
"While these mitigations significantly reduce risk, ChatGPT agent’s expanded tools and broader user reach mean its overall risk profile is higher."
The company noted it also had put work into protecting the agent from adversarial manipulation through prompt injection, which would allow hackers to take control of the tool or manipulate its actions.
"For example, a malicious prompt hidden in a webpage, such as in invisible elements or metadata, could trick the agent into taking unintended actions, like sharing private data from a connector with the attacker, or taking a harmful action on a site the user has logged into," OpenAI added.
"Because ChatGPT agent can take direct actions, successful attacks can have greater impact and pose higher risks."
All web inputs are private, the company revealed, meaning it does not collect or store any data such as passwords.
Make sure to follow ITPro on Google News to keep tabs on all our latest news, analysis, and reviews.
MORE FROM ITPRO
-
Can we disrupt the data center designs?In-depth Distributed approaches or new power sources could enable data centers to grow with minimal harm to the environment
-
Surging memory costs are scuppering digital transformation projectsNews Most organizations are already affected by memory cost increases, with no light at the end of the tunnel any time soon
-
‘AI is not making IT simpler – it's making it more consequential’: IT workers are feeling the heat as AI raises expectationsNews A SolarWinds survey suggests AI makes IT work more strategic, but also adds friction and raises expectations
-
‘Fragmentation is poison’: How Microsoft is targeting disparate data to boost AI adoptionNews Amir Netz, the co-creator of Microsoft's Power BI service, tells ITPro that business context is key to effective AI deployment.
-
Salesforce ramps up agentic AI research with new foundry projectNews Researchers are already working on new tools for agent-to-agent interaction and “ambient intelligence”
-
AI adoption rates aren’t matching IT hypeNews The appetite for AI is there, but a range of issues are hampering adoption
-
Oracle announces new proactive enterprise agents at AI World Tour LondonNews With a slew of new tools and customization options, Oracle is aiming to ground AI agents directly in enterprise data
-
Meta engineer trusted advice from an AI agent, ended up exposing user dataNews The internal security incident exposed sensitive user data to unauthorized employees
-
OpenAI says AI tools are paying dividends for small businesses, but uptake is sluggish in several UK regionsNews While some small businesses are seeing big benefits, many don't use AI at all
-
Microsoft has a new AI poster child in Anthropic – and it’s about timeOpinion Microsoft is cosying up to Anthropic at a crucial time in the race to deliver on AI promises
