ChatGPT Agent does more than announced: discovered functions reveal greater power and fewer controls.
- Graziano Stefanelli
- Jul 30
- 4 min read
Early users have uncovered capabilities that were not officially listed, suggesting that OpenAI’s agent is not just smart—it’s surprisingly autonomous and operationally aggressive.
In the days following the July 2025 rollout of ChatGPT Agent, users across Plus, Pro, and Team tiers began exploring what this assistant could actually do in real-world scenarios. While OpenAI had already teased powerful automation tools—including real web navigation, form filling, and connector integrations—many of the agent’s most surprising behaviors emerged silently from user testing and daily usage.
Not all of them were anticipated. Some go well beyond what was promised. Others raise technical, ethical, and managerial questions. And the most intriguing takeaway from early observations is this: ChatGPT Agent behaves more like a real user than like a bot.
The officially declared features only scratched the surface of the agent's true capabilities.
While OpenAI promised real browsing, form interactions, and connector-based actions, testers uncovered functions that behave as if the agent were using a real mouse and memory buffer.
OpenAI’s formal release notes for ChatGPT Agent described a system capable of performing real-time browsing, interacting with web pages, reading and responding to emails, and working with files in Google Drive or GitHub—all under explicit user permission. These baseline promises have been confirmed. But they’re only the beginning.
Over the past ten days, users have demonstrated that the agent can:
click on CAPTCHA-style “I’m not a robot” checkboxes autonomously;
download public files such as PDFs and CSVs;
analyze table data from websites in a structured way;
generate, store, and later reuse intermediate files, like CSVs pulled from scraped content;
sort folders and contents inside Google Drive, not just read from it;
compose and send real emails via Gmail, based on user-approved connectors;
interact with JavaScript-heavy websites, even completing booking or checkout steps—though with limitations.
Some of these behaviors were never explicitly announced. In particular, the agent’s capacity to store temporary outputs across multiple actions, open multiple browser tabs, or click through basic CAPTCHA elements was discovered through user experiments, not technical documentation.
Discovered actions suggest a semi-autonomous behavior chain that wasn't disclosed.
Even though OpenAI framed the agent as "user-guided," its actual behavior includes persistent context, parallel web navigation, and compositional workflows.
One of the most notable discoveries is the agent's short-term operational memory. Unlike standard prompts, where context resets quickly, ChatGPT Agent can:
remember and reference files it created 10–15 actions earlier;
pull intermediate results from websites and reuse them across services;
maintain a consistent execution plan across multi-step tasks such as: search → compare → extract → import → notify.
Users have reported that when asked to find an Airbnb listing, the agent not only compiles a list of suitable candidates, but also stores price and availability data in a structured format, downloads a CSV, and prepares a slide presentation—all in sequence. This behavior was not highlighted in the official onboarding nor reflected in the current documentation.
Even more unexpected: some testers discovered that the agent creates transient browser tabs in parallel, particularly when executing comparison tasks. Though not always stable, this behavior hints at a more complex internal browser orchestration, something that no other consumer AI currently offers.
Certain promised features are still missing or unavailable to regular users.
Despite exceeding expectations in some areas, ChatGPT Agent lacks transparency, audit controls, and granular role restrictions—especially in Plus and Pro accounts.
While the agent surprised users with its autonomy, other expected enterprise-grade capabilities are still missing from the experience. These include:
granular permissions: currently, Gmail or Drive connectors require full access; there’s no support for “read-only” or folder-specific scopes;
auditing dashboards: there is no visual log of what the agent did during a session unless the user was watching in real time;
"Watch mode" or passive simulation: users cannot yet run a task in observation mode to preview actions without execution;
integration with other apps like Slack or Notion: promised in Enterprise tiers, these are not available to the public;
BYO connector or API integrations: there is no UI or API support for custom third-party connectors.
OpenAI has emphasized that many of these controls are planned for Team Enterprise deployments, not for individual users. But this creates a paradox: the most autonomous version of the agent is currently in the hands of unsupervised users, not behind corporate governance walls.
The current behavior of ChatGPT Agent feels more powerful and less constrained than officially anticipated.
The assistant is not just executing steps—it is composing logic, manipulating apps, storing partial results, and performing tasks as if it were a human operator.
The most consistent feedback from early testers is that ChatGPT Agent acts more like a digital worker than a prompt responder. It doesn’t just return answers—it interacts with the world, navigates real websites, creates artifacts, clicks through workflows, and moves from tab to tab like an assistant with a screen, mouse, and memory.
And while this opens up tremendous creative, operational, and productivity potential, it also exposes several risks:
users can't fully track what actions were taken once the agent finishes a session;
the agent can inherit credentials and permissions, which introduces long-tail security risks;
it performs actions faster and more aggressively than a human would, triggering anti-bot systems (and sometimes bypassing them).
In some cases, these capabilities deliver magic. In others, they raise legitimate concerns.
____________ FOLLOW US FOR MORE.
DATA STUDIOS
