ChatGPT Agent Mode: AI on the Web
A Different Approach: Cloud Instead of Desktop
Section titled “A Different Approach: Cloud Instead of Desktop”Claude Cowork works on your computer — with your local files. ChatGPT Agent Mode OpenAI's agentic feature in ChatGPT that combines a visual browser, code execution, file editing, and app integrations. The agent works in the cloud and can independently execute multi-step tasks on the web. works in the cloud — with its own virtual browser that can navigate websites, fill out forms, and complete online tasks.
Both solve the same problem: AI that doesn’t just answer, but acts. But their strengths lie in different places.
What Agent Mode Can Do
Section titled “What Agent Mode Can Do”Agent Mode unifies three capabilities that OpenAI originally developed separately:
| Component | What It Does | Originally |
|---|---|---|
| Visual Browser | Navigates websites, clicks, scrolls, fills forms | Operator (Jan 2025) |
| Deep Research | Searches the web for 5–30 minutes, analyzes hundreds of sources | Deep Research (Feb 2025) |
| ChatGPT Core | Conversation, code execution, file editing | ChatGPT |
Toolkit
Section titled “Toolkit”| Tool | Function |
|---|---|
| Visual Browser | Interacts with websites via GUI (clicking, scrolling, typing) |
| Text Browser | Simpler, reasoning-based web queries |
| Terminal | Code execution (Python), data analysis |
| File System | Create spreadsheets, presentations, documents |
| Apps | Gmail, Google Calendar, Drive, GitHub, Slack, Notion, HubSpot, and more |
Practical Examples
Section titled “Practical Examples”- “Check my calendar and brief me on upcoming meetings based on recent news about the participants.”
- “Research the three largest competitors in our market and create a comparison slide deck.”
- “Find the cheapest flights to Berlin next week and create a comparison table.”
Deep Research: The Research Assistant
Section titled “Deep Research: The Research Assistant”Deep Research deserves special attention because it solves a task knowledge workers face daily: Comprehensive research with sources.
- Independently searches the web for 5–30 minutes
- Analyzes hundreds of sources
- Creates structured reports with citations
- Can narrow scope to specific websites
- Can work with uploaded files
| Plan | Deep Research Queries/Month |
|---|---|
| Free | 5 |
| Plus ($20) | 25 |
| Pro ($200) | 250 |
Limitation: Occasional hallucinations. Always spot-check sources.
Scheduled Tasks
Section titled “Scheduled Tasks”ChatGPT can plan recurring tasks:
- One-time, daily, weekly, monthly, or yearly
- Maximum: 10 active tasks at once
- Works on web, desktop, and mobile
Examples:
- “Every Friday: Weekend plan based on weather and location”
- “Every Monday: Summary of my GitHub activity”
- “Daily at 7 AM: Industry news briefing”
Pricing
Section titled “Pricing”| Plan | Price | Agent Mode | Deep Research |
|---|---|---|---|
| Free | $0 | No | 5/month |
| Plus | $20/month | Yes (40 msg/month) | 25/month |
| Pro | $200/month | Yes (400 msg/month) | 250/month |
| Team | $25–30/user/month | Yes | 25/month |
Important: Only initial requests count toward the quota. Follow-up questions and clarifications within a task don’t.
Strengths and Limits
Section titled “Strengths and Limits”Strengths
Section titled “Strengths”- Web automation — the biggest difference from Cowork: Agent Mode can navigate websites, fill forms, conduct interactive web research
- Cross-platform — web, desktop (Mac + Windows), iOS, Android
- Deep Research — one of the best available solutions for comprehensive web research
- App integrations — Gmail, Calendar, Drive, Slack, Notion, GitHub, and many more
Limits
Section titled “Limits”- No local file access — everything runs in the cloud. You can upload files, but Agent Mode can’t access your local folders
- Anti-bot blocks — many websites (Amazon, LinkedIn, social media) block the visual browser
- CAPTCHAs — needs your help with “I’m not a robot”
- Speed — multi-step tasks take 1–10 minutes as each page is processed individually
- Complex authentication — MFA, SSO redirects, and OAuth are problematic
- Audit gap — agent actions are indistinguishable from user actions in logs
Comparison: Cowork vs. Agent Mode
Section titled “Comparison: Cowork vs. Agent Mode”| Aspect | Claude Cowork | ChatGPT Agent Mode |
|---|---|---|
| Focus | Desktop and local files | Web and cloud |
| Local file access | Yes (folder sandbox) | No |
| Web browsing | Chrome extension | Dedicated visual browser |
| Platforms | macOS + Windows desktop | Web, desktop, iOS, Android |
| Sub-agents | Yes (parallel) | No |
| Deep Research | Web search with summarization | 5–30 min autonomous research |
| Scheduled tasks | Yes (requires app open) | Yes (up to 10 active) |
| Entry price | $20/month | $20/month |
| Best for | Document creation, file work | Web automation, research |
When to Choose Which?
Section titled “When to Choose Which?”Cowork when you:
- Work with local files (organizing PDFs, Excel analysis, presentations)
- Need cross-app workflows (Excel → PowerPoint)
- Primarily work at your desktop
Agent Mode when you:
- Have web-based tasks (research, bookings, forms)
- Need Deep Research for comprehensive analysis
- Want to work mobile or cross-platform
Both when you have different task types — which is the case for most knowledge workers.
Try It Yourself
Section titled “Try It Yourself”Exercise 1: Test Agent Mode
Section titled “Exercise 1: Test Agent Mode”Activate Agent Mode in ChatGPT (Tools dropdown or /agent). Set a multi-step web task: “Find the 3 most-cited studies on [your field] from 2025 and summarize each in 3 sentences.”
Exercise 2: Use Deep Research
Section titled “Exercise 2: Use Deep Research”Start a Deep Research query on a topic you need to research anyway. Compare the output with what you would have found in 30 minutes on your own.
Exercise 3: Comparison Test
Section titled “Exercise 3: Comparison Test”Take a task and complete it with both Cowork and Agent Mode. Document: Which was faster? Which produced higher quality? Which was more convenient?
Looking Ahead
Section titled “Looking Ahead”Claude Cowork and ChatGPT Agent Mode represent two philosophies: desktop-first vs. cloud-first. Both are improving, both will converge. For you as a knowledge worker, the most important skill isn’t mastering one tool perfectly — it’s knowing when each is the right choice.
In the next lesson, you’ll tackle perhaps the most important skill when working with AI agents: Trust Calibration — when you can trust, when you must verify, and how to calibrate that balance over time.