Last week I gave a ChatGPT agent a straightforward job: find three competitors’ pricing pages, pull the tiers into a table, and flag anything that looked new since January. It took eleven minutes. Six of those were spent clicking through cookie banners and a “request a demo” wall it never quite got past. The other five produced a comparison table I actually used in a meeting.
That ratio, roughly half wasted motion, is the honest state of things right now. Agent mode is genuinely capable and genuinely clumsy in the same afternoon. Working out which tasks land on the right side of that line is most of the skill.
What OpenAI actually means by “agent”
Regular ChatGPT answers questions. An agent acts. In agent mode the model gets its own sandboxed computer: a browser it drives, a terminal, and a small file system. It can open a site, scroll, click, type into a form, download a PDF, parse it with Python, build a spreadsheet, and hand you the finished file at the end.
OpenAI shipped this in July 2025 for Plus, Pro and Team subscribers. The same capabilities later leaked onto phones, and the result is a mobile experience where you can hand off a task and watch it work while you’re on a train. That shift is covered in more detail in the ChatGPT mobile app’s new voice-based agentic features, and it’s a bigger change than it sounds. Talking to something that then goes and does the clicking is a different relationship than typing into a box.
What a run actually looks like
You describe a goal. A panel opens on the right and the agent starts narrating its own actions: opening a page, scrolling, clicking a link, running a snippet of code. You can watch the whole thing, stop it mid-step, or grab the virtual browser yourself when it gets stuck on something only a human can spot.
Approval gates and the sandbox
Anything sensitive triggers a confirmation. Logging into an account, spending money, sending an email, deleting a file. The sandbox also blocks a lot by default, which is why the agent will sometimes announce it can’t reach a site that loads fine in your own browser. Annoying, but it’s the right default while the failure modes are still being mapped.
The clock is real
A single run has a time budget, and long research chains hit it. I’ve had agents get 80% of the way through a five-source literature sweep and then stop with a polite note about the limit. The work up to that point is usually intact, so short, staged tasks beat one heroic request.
Jobs it handles well
Anything repetitive, browser-shaped, and tolerant of small errors is a good fit. In practice that means:
- Form-filling at volume. Twelve supplier quotes into one spreadsheet, with source URLs captured next to each figure.
- Competitive monitoring. Checking a set of pricing and changelog pages and reporting only what moved.
- Data cleanup. Taking a messy CSV export, deduplicating it, and returning something a pivot table won’t choke on.
- Screenshot and file assembly. Grabbing a set of images, resizing them, and dropping them into a folder structure you specified.
- First-pass meeting prep. Pulling recent news on a company and summarising it into a one-pager, with links you can check.
That last one is where agents start overlapping with the dedicated research tools, and the quality gap between them is wider than the marketing suggests. If deep, citation-heavy work is your use case, it’s worth reading a comparison of the best AI agents for research and how to tell which ones are making things up before you commit an afternoon to one.
Where it falls apart
Three failure modes show up over and over.
It believes the page. A site with stale pricing will be reported as current. A blog post from 2021 gets cited as if it were news. The agent is good at reading and bad at doubting.
It gets stuck politely. CAPTCHAs, two-factor prompts, and multi-step logins will stall a run for ten minutes before it tells you. Watch the first few minutes of any new task type and you’ll spot the pattern instantly.
It produces confident-looking artefacts. A slide deck with a broken chart axis. A summary that quietly merges two different companies’ numbers. The output looks finished, which is exactly what makes it dangerous if nobody reads it.
Prompting an agent is not prompting a chatbot
With a chatbot, a short question gets a short answer. With an agent, a short request gets a long chain of decisions you didn’t specify. Four things belong in almost every brief:
- The goal, stated as an outcome. “A single CSV with columns X, Y, Z” beats “look into pricing.”
- The constraints. Which sites, which date range, which sources are off-limits.
- A definition of done. “Stop when you have twelve rows or you’ve checked all nine URLs.”
- Checkpoints. “Show me the table before you build the report.”
Tell it to cite the URL next to every extracted number, too. It won’t always comply, but the request changes how carefully it reads.
Agent mode is not the only agent you have
It’s easy to forget that ChatGPT’s own ecosystem has several overlapping things called agents. Custom GPTs with actions call APIs on your behalf. Connectors let the model read your Drive or Slack without any browser at all. Agent mode is the general-purpose clicking one, and it’s the slowest and least reliable of the three when a direct integration exists.
The same pattern holds outside OpenAI. Developer-focused coding agents have gotten very good at narrow tasks and remain shaky at anything requiring judgement about a codebase’s conventions. There’s a useful breakdown of the best AI agents for developers, judged by what they get wrong that takes that critical angle seriously rather than benchmarking happy paths.
For teams running these things in production rather than on a laptop, the questions change completely: rate limits, audit logs, permission scoping, and what happens when an agent is 90% right at 2am. That’s the focus of which enterprise AI agents hold up under real production load, and it’s a sobering read if you’re planning a rollout.
Handling logins without handing over your life
The temptation is to give the agent your real credentials so it can log in and get on with it. Resist that on anything tied to money or customer data. Better options, roughly in order of preference: use an export instead of a login, create a read-only account with a scoped password, or use a password manager’s one-time share link that expires in an hour.
Two other habits pay off. Keep a separate browser profile for agent work so a stray click can’t reach your saved sessions. And read the run log after any task that touched an account, because the narration is the only record you’ll get of what was actually visited.
The skill worth building now
Supervising an agent is turning into a distinct competence, and it looks less like writing prompts than like managing a fast, literal junior colleague. You scope the work tightly. You define what finished looks like before you start. You check the first ten minutes rather than the last. You read the output with suspicion, especially when it’s neat.
Start with one boring, recurring task you already know how to do by hand. Run the agent on it while you do something else, then compare its output against your own. Do that three weeks running and you’ll have a decent feel for where it’s dependable and where it’s bluffing, which is worth more than any list of features. The people getting real leverage out of ChatGPT agent right now aren’t the ones giving it the biggest jobs. They’re the ones who’ve gotten good at spotting the exact moment it stops being useful.

