Product update: OpenAI says ChatGPT agent is no longer available and directs people to ChatGPT Work for longer, multi-step tasks and finished deliverables. The examples below address the kinds of tasks readers may have searched for as “Agent Mode” examples, but they are not promises that the discontinued Agent Mode interface is still available. (OpenAI Help Center)
To try a Work task, open ChatGPT on web or mobile and start a task in Work. Availability may depend on your plan, workspace settings, and rollout access; in a managed workspace, ask your administrator whether Work is enabled for you. (ChatGPT Work and Codex) If a task needs the cloud browser, that feature is available on paid ChatGPT plans in supported regions, excluding Free and Go; availability can also depend on workspace permissions and rollout. (Using cloud browser in ChatGPT)
Quick answer: examples of agent-style tasks
Use an agent-style workflow when you want help carrying a task through several steps—not just getting a quick answer. OpenAI’s published examples for ChatGPT Work include analyzing a month-end budget variance, turning source material into a marketing brief, and preparing for a sales meeting. Its announcement also describes creating materials such as spreadsheets, slides, documents, and web apps from information across apps and workflows; which connected apps and actions are available can depend on permissions and workspace settings. The prompts below are illustrative drafts, not tested prompts. Supply relevant files or context, check the output, and approve consequential actions yourself. (ChatGPT Work announcement)
| Goal | Adaptable prompt | Provide | Review before using |
|---|---|---|---|
| Analyze a budget variance | “Compare these two monthly budget files. Summarize the largest changes by category, show the figures you used, flag anything you can’t reconcile, and prepare a short report. Don’t change the source files.” | The two reports, period labels, and any definitions for categories | Check totals and whether comparisons use the right periods; verify any unexplained discrepancies. |
| Turn research into a campaign brief | “Use these customer research notes and product facts to draft a campaign brief. Separate evidence from suggestions, note missing information, and create a version for each audience listed. Don’t publish or send anything.” | Research, approved product facts, audience descriptions, and brand guidance | Confirm claims against source material; have a person approve the brief and any assets before publication. |
| Prepare for a sales meeting | “Use these call notes and account materials to create a meeting brief: account context, open questions, likely priorities, and a draft follow-up email. Mark uncertain details and do not send the email.” | Call notes and authorized account information | Verify customer details, distinguish facts from inferences, and approve the email before sending. |
| Compare competitors and draft a presentation | “Compare these three competitors using their public product pages. Cite the source for each factual claim, note when information is missing, and draft a slide outline. Don’t contact anyone or make purchases.” | Competitor names or URLs and the comparison criteria | Recheck important claims at their sources and confirm that the comparison is current before presenting it. |
| Build a repeatable team workflow | “For each new software request, summarize the request, compare it with this policy, identify missing information, and draft the approval or rejection rationale. Ask a person to decide; don’t open a ticket or send a response.” | The request, policy, and approved criteria | Confirm policy interpretation and have an authorized person make the decision. |
The first three examples adapt tasks OpenAI has described for ChatGPT Work. The competitor-analysis and presentation example adapts a task from the earlier ChatGPT agent launch; the team workflow is illustrative, not a verified configuration or outcome. (ChatGPT Work announcement; ChatGPT agent launch)
How to choose: chat, ChatGPT Work, or an API
- Use ordinary chat when the task is mainly to explain, rewrite, brainstorm, or analyze material you provide and you do not need a workflow across tools. This is a practical rule of thumb, not a formal product boundary.
- Try ChatGPT Work for an individual multi-step task—such as preparing the sales-meeting brief above—when Work is available to your account and the needed files or app connections are permitted. OpenAI describes Work as designed for longer tasks and finished deliverables. Workspace settings, app access, and rollout can affect what you can do. (ChatGPT Work and Codex)
- Build with an API when the workflow belongs inside your own application or service and you need to define its integrations, permissions, review gates, and behavior. The ChatGPT experience and an API integration are different products; do not assume an API call reproduces ChatGPT Work. OpenAI recommends the Responses API for new API projects. (Migrate to the Responses API)
For example, preparing one sales-meeting brief for yourself may fit a ChatGPT Work task. If your product needs to prepare a brief whenever a customer record changes, an API workflow may be more appropriate: your service supplies authorized account data, asks the model to draft a brief, and keeps any email-sending step behind a person’s approval.
Translating a sales brief into a production API workflow
OpenAI recommends the Responses API for new API projects. For this sales-brief example, one possible application design is to use it for the model interaction and have your application define narrowly scoped tools for retrieving approved account data or saving a draft. Treat those tools and review steps as design choices for your application, not as documented capabilities of ChatGPT Work. (Migrate to the Responses API)
A practical sequence is:
- Define the job: provide the account context and ask for a brief with factual claims, open questions, and a clearly labeled draft follow-up.
- Limit access: expose only the account fields and actions this workflow needs. For a deployed workload, OpenAI recommends using a project service account and its API key. (API keys and service accounts)
- Let the application control actions: the model can request a tool call, but your code should validate the request, enforce your authorization rules, and decide whether to execute it. Do not make email sending automatic unless your approval policy explicitly allows it.
- Review and record: show the draft to an authorized person before sending; keep appropriate records of the input, draft, approval, and action.
This is an implementation outline, not a drop-in recipe. Confirm the model, tool definitions, input schema, authentication setup, and response handling against the current API documentation before implementation. The ChatGPT Agent Mode interface is not presented here as directly callable through the API.
Safety checks for examples involving apps or external actions
Review outputs and external actions before relying on them, especially when they affect customers, money, records, or access. Treat these as editorial recommendations—not claims that ChatGPT Work enforces a particular approval step. App permissions and availability can depend on the connected account, app, and workspace. Older Agent Mode guidance about confirmations or “watch mode” concerned that discontinued feature; this article does not assume those safeguards apply to ChatGPT Work. (Managing app permissions in ChatGPT)
Bottom line: start with a specific deliverable—such as a budget analysis, campaign brief, sales preparation, or competitor presentation. For an individual multi-step task, open Work in ChatGPT if it is available to your account; ask your workspace administrator if it is not visible in a managed workspace. Use an API when you need to embed and govern a workflow in your own product. Treat prompts as starting points, verify the output, and require human approval for consequential actions.
Sources
ChatGPT agent — OpenAI Help Center
ChatGPT Work and Codex — OpenAI Help Center
ChatGPT is now a partner for your most ambitious work — OpenAI
Introducing ChatGPT agent — OpenAI
Using cloud browser in ChatGPT — OpenAI Help Center
Managing app permissions in ChatGPT — OpenAI Help Center