ChatGPT: Expert

Expert

The expert move in ChatGPT is to stop asking and start delegating: handing a multi-step task to the agent and supervising it instead of doing each step yourself. It is genuinely autonomous, and it is supervised by design. This is a reliable workflow applied to an autonomous run: scope tight, watch, verify.

Check the Swiss caveat first

Confirm availability before you rely on it

As of July 2026, OpenAI’s own documentation is inconsistent about the European Economic Area and Switzerland: some pages say the agent is available in all supported countries, others still carry EEA/Switzerland/UK exclusions, and the rollout has been uneven in practice. A Swiss workspace may not have the agent yet. Confirm it is present on the actual target account before promising it to a client.

The agent is a paid feature (Pro, Plus, Business, Enterprise, and Edu). Usage is capped per month; the widely cited launch figures were roughly 40 agent messages on Plus and around 400 on Pro, but OpenAI adjusts limits without much notice. Verify the current cap in the app or OpenAI’s help center before planning work around it.

What the agent actually does

Launch it from the composer’s tools menu and it spins up a sandboxed virtual computer with its own browser and file tools. It runs the loop on its own (gather context, act, observe), see How AI coding tools work:

• Browses sites, clicks, reads, and fills and submits forms
• Opens files you uploaded and builds or edits spreadsheets and slides
• Pulls from your connected Apps

You steer from a live activity view and can interrupt or take over the browser at any time. Runs are session-based and usually last about 5 to 30 minutes. Nothing runs unless you launch it; there is no always-on listener.

Tip

Treat the agent as a supervised intern. Scope the task narrowly, watch the activity view, and rely on the approval prompts: it pauses for permission before consequential actions like sending an email or making a purchase.

A good agent run

1. Plan first. Run deep research (a read-only, cited research pass) or a Thinking-model chat to nail the plan before anything acts.
2. Scope tight. Give the agent one clear objective and the inputs it needs, not an open-ended goal.
3. Watch and approve. Follow the activity view; approve consequential steps deliberately.
4. Verify the output. Open the cited sources; treat anything ungrounded as a draft.

Automate the recurring work

Surface	What it’s for
Scheduled tasks	Recurring or monitoring jobs and reminders that run on a cadence and notify you; the closest thing to cron. Available on Plus, Pro, Business, and Enterprise.
Agent	Interactive, multi-step automation, but session-based, not a daemon. It will not watch an inbox on its own.
Shared GPTs and Projects	The reusable-asset story for a team: package a workflow once, share it.

Caution

For sensitive accounts, know about Lockdown Mode. It blocks the exfiltration paths prompt-injection attacks rely on by disabling live web browsing, web images, deep research, and the agent itself. That makes it a control for sensitive work, not a companion to agent runs: with Lockdown on, the agent is off. OpenAI notes it reduces, but does not eliminate, the risk.

Where the consumer app stops

The ChatGPT app is not a scripting surface. There is no consumer CLI or CI hook, and the agent is interactive rather than headless. True headless automation, parallel multi-agent orchestration, and anything you’d wire into a pipeline belong to the OpenAI API and Codex: a different tool, covered in its own track.

Note

AI is an accelerator with no opinion about direction. The agent will pursue exactly the goal you scope, which is why tight scoping and verification are the whole job at this level.

You’re done when…

• ✅ You’ve confirmed whether the agent is enabled on your account
• ✅ You’ve run an agent task end to end, supervising it
• ✅ You’ve set up a Scheduled task
• ✅ You know which work belongs in the API/Codex instead

Next step

You’ve reached the top of the ChatGPT track. Pick another tool, or revisit the ideas underneath all of them.

Foundations The ideas underneath every tool in these docs.

Start here Pick another tool track.