ChatGPT used to answer questions. In 2026, it does your work. Agent Mode is OpenAI's autonomous feature that lets ChatGPT open a browser, click buttons, fill forms, run code, read your files, and hand you back a finished result — not a plan, the actual thing. This chatgpt agent mode guide walks through what it does, when to use it, and five prompts that produce something useful on the first try.
What Agent Mode Actually Is
Agent Mode is a runtime, not a model. When you flip it on, ChatGPT spins up a sandboxed virtual machine with a real browser, a terminal, and a filesystem. It can navigate to a URL, log in if you gave it credentials, scrape a table, download a PDF, run Python on the result, and write a file back to your Drive or your machine. You watch it work in a side panel and can pause it at any point.
The important shift: earlier ChatGPT features like browsing or Advanced Data Analysis were single tools. Agent Mode chains them together and makes decisions between steps. If step 3 fails, it retries or picks a different route without asking. That's the “agent” part — it has a goal and it iterates.
How to Turn Agent Mode On
Open ChatGPT on web or desktop. In the message composer, click the tools icon (the small slider) and pick Agent. You'll see a new side panel appear — that's the sandbox. Type your goal in natural language and hit send.
A few things to know before you fire:
- Approvals: the agent pauses before real-world consequences — payments, sending an email, submitting a form. You confirm each one.
- Login: for auth-gated sites, you take over the browser panel, log in, then hand control back. Credentials never leave the session.
- Time budget: most tasks finish in 5–20 minutes. Long research jobs can run 40+.
- Files: upload at the start, or connect Google Drive, GitHub, Notion, or SharePoint through Connectors.
Five Agent Mode Prompts That Actually Work
Vague prompts (“research competitors”) waste the agent's runtime. Specific prompts with a goal, a source, and an output format finish cleanly. Here are five patterns worth stealing.
1. Competitive pricing sweep.
Go to the pricing page of each of these SaaS tools: Notion, Airtable, Coda, ClickUp, Monday.
For each, extract: plan name, monthly price per seat (USD, annual billing), user cap, and the single headline feature that plan unlocks.
Return a single CSV I can download, with a note flagging any plan whose price changed vs. the archive.org snapshot from 6 months ago.
2. Inbox-to-CRM sync.
Open Gmail via the Google connector. Find every email from the last 14 days where the sender domain is NOT in this allowlist: [gmail.com, our own domain, known vendors list attached].
For each unique sender, extract: name, company, what they want, and a suggested next action. Skip newsletters and calendar invites.
Output as a Google Sheet in my Drive called 'Inbound leads — this fortnight'.
3. Content refresh audit.
Crawl every blog post on promptspace.in/blog published before 2026-01-01. For each, check: is the H1 keyword still ranking on Google page 1 for a UK IP? Are any external links broken? Is the featured image still loading?
Return a table with URL, three health checks, and a one-line 'refresh priority' verdict (high/medium/skip).
4. Travel booking, held for approval.
Find me a return flight Delhi to Lisbon, departing Nov 12, returning Nov 22. Non-stop preferred, one stop OK if it saves >30%. Aisle seat, one checked bag. Budget: under 65,000 INR.
Compare Google Flights, Skyscanner, and the airline direct sites. Pre-fill the booking on the cheapest option and pause for my approval before payment.
5. Repo triage.
Open the GitHub repo promptspace/api. List every open issue with no assignee older than 14 days. For each, read the thread and classify it as: bug / feature / question / stale.
Post a comment on each stale one asking if it's still relevant, and label all bugs with 'needs-repro'. Give me a summary table at the end.
Agent Mode vs. Deep Research vs. Custom GPTs vs. Regular ChatGPT
The four options overlap, which is why people pick the wrong one. Here's the honest split:
| Feature | Regular ChatGPT | Custom GPT | Deep Research | Agent Mode |
|---|---|---|---|---|
| Answers questions | Yes | Yes | Yes, cited | Yes |
| Browses live web | Limited | Limited | Yes, read-only | Yes, read + click |
| Clicks buttons, fills forms | No | No | No | Yes |
| Runs code on your files | Yes (ADA) | Yes | No | Yes |
| Chains 10+ steps autonomously | No | No | Yes | Yes |
| Typical run time | Seconds | Seconds | 5–30 min | 5–40 min |
| Best for | Quick answers | Repeatable workflows | Deep reports | Doing real work |
Rule of thumb: if the task ends in a file, a booking, a form submission, or a status change somewhere — reach for Agent Mode. If it ends in a document you read — Deep Research is faster and cheaper.
Where It Still Falls Over
Two years in, Agent Mode is powerful but not magic. It struggles with:
- CAPTCHAs and aggressive bot detection. It'll pause and ask you to solve them.
- Heavy JavaScript SPAs where selectors change on every load — some banking portals.
- Ambiguous goals. “Make my landing page better” produces something you won't like. Specify the metric.
- Cost-blind decisions. It doesn't know your budget unless you say so. Cap the spend.
- Multi-account confusion. If you're logged into two Gmails, tell it which one.
Don't hand it credentials for anything you can't afford to lose. Use a scoped API key or a burner login. OpenAI's official Agent Mode documentation is worth reading before you connect anything critical.
FAQ
Is Agent Mode available on the free plan?
No. It requires Plus, Pro, Business, or Enterprise. Free users can watch demos but can't run tasks.
How much does a run cost?
Runs are metered against a monthly quota — around 40 runs on Plus, 400 on Pro, higher on Business. Overages are billed per task.
Can it access sites behind my company VPN?
Only if you use the desktop app with the Local Bridge feature enabled. The cloud sandbox can't reach private networks.
Does it remember previous runs?
Each Agent Mode task starts in a fresh sandbox. Use Memory or a Custom GPT to persist context across runs.
Can I run two agents in parallel?
Yes on Pro and above. Plus is single-threaded.
Is my data used for training?
No, Agent Mode sessions are excluded from training on all paid tiers by default.
Can it work with my Notion/Slack/Jira?
Yes via Connectors — you approve the scope, and the agent gets read/write access limited to that scope.
What happens if it makes a mistake?
You can pause, rewind to any step, edit the plan, and resume. Every action is logged.
Where to Go Next
Agent Mode rewards people who write like managers, not like Googlers. State the goal, name the source, define “done,” and specify the output. The prompts above are a starting kit — bend them to your workflow, save the winners, and share the losers so nobody else runs them.
If you want a head start, we curate prompts specifically tuned for autonomous agents on our free prompt library, and we break down more workflows like this one on the PromptSpace blog.
Browse 5,000+ ready-to-use prompts at PromptSpace.
