Agents · Lesson 01 Pro ~12 min read Browser + actions

ChatGPT Agent: real autonomy in your browser.

ChatGPT Agent is the agentic mode that absorbed Operator and Deep Research into one unified system. It controls a virtual browser, fills forms, navigates websites, runs research, and delivers finished work. Different category from chatting with ChatGPT — you delegate a task, then come back to a result.

The mental model

Agent ≠ Chat. You're delegating a task, not having a conversation.

Regular ChatGPT works in turns — you ask, it answers, you refine. ChatGPT Agent works in tasks — you describe an outcome, it works on it (sometimes for 5-30 minutes), and reports back when done. You can interrupt to redirect, but you're not chatting through every decision.

The agent has a real virtual browser. It can log into websites you authorize, fill in shopping carts, comparison-shop across multiple sites, book reservations, scrape data into spreadsheets. It's basically a remote worker with internet access.

What it can't do: anything you haven't authorized, payments without confirmation, account creation on your behalf, anything illegal, and it asks before high-stakes actions like 'place this $400 order.'

Workflow 01 Multi-site research with synthesis

1

Comparison shopping or vendor research

The original Operator pattern — give the agent a task that spans multiple sites, then come back to a synthesized result.

The prompt that works

Multi-site comparisonI need to find the best wireless conference-room speakerphone for under $500. Compare options across: • Best Buy • Amazon • B&H Photo • Manufacturer sites for Jabra, Logitech, Poly, and Yealink For each top option, give me: • Model name and current price across all sources • Pros from professional reviews (cite the review source) • Common complaints from user reviews • Battery life and connection type (Bluetooth/USB-C/dongle) Present a final ranked recommendation table with one-sentence rationale. Don't buy anything — just produce the comparison.

Best use cases

  • B2B vendor evaluation
  • Personal purchase research
  • Competitive analysis (your competitors' websites, pricing pages)
  • Travel comparison across booking sites
Always tell the agent NOT to purchase. The default behavior is to ask before buying, but explicit instruction is safer.
Time savings: An afternoon of comparison shopping → 15-30 min of agent time.

Workflow 02 Form-filling and data entry at scale

2

When you have repetitive web forms to complete

Government forms, vendor onboarding pages, registration flows — anything where 'fill in the same info 30 times' is the task. The agent can plow through it.

The prompt that works

Batch form-fillingI have a list of 25 vendors that need to be added to our supplier portal (vendor-list.csv attached). For each vendor: 1. Open the portal at supplier-portal.example.com 2. Click 'Add New Vendor' 3. Fill in fields: Company name, contact email, tax ID, address — all from the CSV 4. Upload their W-9 form from the linked Google Drive 5. Submit and capture the confirmation number Produce a results CSV with: vendor name, status (success/failed), confirmation number, any errors. Stop and ask if you hit a CAPTCHA or weird form behavior.

Best use cases

  • Supplier or contact data entry at scale
  • Account creation across multiple SaaS tools (your tools, not on behalf of others)
  • Government compliance form batch submissions
  • Survey or registration form completion
CAPTCHAs are deliberately designed to block agents. The agent will stop and ask you to solve them. Plan for this if your target site has them.
Time savings: 25 forms × 5 min each = 2 hours manual → ~25 min agent supervised.

Workflow 03 Research → write → publish workflow

3

End-to-end content production from a brief

The Deep Research + Operator combo lets you go from prompt to published draft in one task. Research a topic, write the content, optionally schedule it.

The prompt that works

Research → draft pipelineWrite a 1,500-word blog post on 'How AI is changing customer support in B2B SaaS.' Research phase: Find 10-15 current sources (last 6 months). Mix of analyst reports, vendor case studies, and reporter coverage. Cite sources inline as you write. Writing phase: Hook with a specific anecdote (not abstract intro). Three main sections: what's changing, who's doing it well, what to avoid. End with concrete advice for someone running B2B support today. Output: deliver as a markdown file. Don't publish anywhere — just produce the draft. Add a 'Sources' section at the bottom with all 15 references.

Best use cases

  • Blog drafts with real sources
  • Whitepaper or report production
  • Newsletter content from current events
  • Investor or board memo drafting
ChatGPT Agent can still produce confident-but-wrong claims — fact-check anything before publishing. Treat the output as a strong first draft, not a finished product.
Time savings: A well-researched blog post: 6+ hours manually → 30-45 min agent + your review.
About ChatGPT Agent

Released in July 2025 by OpenAI as the successor that merged Operator (browser-action agent) and Deep Research (long-running research agent). Available to ChatGPT Plus, Pro, and Team users. Free-tier users get limited access.

Pick a multi-hour task and delegate it

Think of a task that would take you 3+ hours manually — research, comparison shopping, data entry, drafting. Hand it to ChatGPT Agent. Don't multitask while it works; just check back every 10 minutes to see if it needs you. Notice the difference between 'I'm doing this' and 'I'm reviewing this.'

Back to ChatGPT Mastery → Try a prompt in the playground →

What you can do now

  • Understand when to reach for Agent vs. regular ChatGPT
  • Specify NOT to make purchases or commitments without explicit approval
  • Use Agent for tasks that span multiple sites or sources
  • Always review the output — it's a strong draft, not finished work
  • Know that CAPTCHAs will pause the agent, by design
Pro
Up next in ChatGPT Mastery

Lesson 02 · Deep Research — long-running multi-source research

When you need depth more than breadth, Deep Research is the better tool. Multi-source synthesis, citation-heavy output, and now connected via MCP to your own tools and data. See the track →