The short answer
Yes, by OpenAI's own definition. The help center states plainly that “Work is an agent designed for longer, multi-step work and finished deliverables” ( help.openai.com, checked 2026-09-14). Straight from the vendor, in the vendor's own words. Argument over.
So “is ChatGPT an AI agent” is the wrong question to keep arguing about. The useful one is narrower: is it enough of an agent for what you actually need, or do its documented limits push you toward a dedicated platform? That's what the rest of this article answers, with checkable numbers, not opinions.
For the deeper theory on what separates a chatbot from an agent, see the deeper chatbot-vs-agent distinction. We stick to ChatGPT here.
What “Agent” Means Here, in One Paragraph (Not a Full Definition Essay)
An agent, in plain terms, is software that takes multi-step action toward a goal without you re-prompting it at every step: it plans, uses outside tools, and keeps going until the job is done. We cover the fundamentals of what makes something an AI agent elsewhere. For a non-coder, the test is simpler: can it act across your other apps, on a schedule or trigger, without you typing the next instruction each time? That's the bar ChatGPT has to clear, and the next section walks through exactly where it clears that bar and where it doesn't.
What ChatGPT Actually Ships Today (as of 2026-09-14, Sourced to OpenAI)
ChatGPT now organizes itself into three modes: Chat, Work, and Codex. OpenAI's help center draws the line: “Chat is for fast, conversational assistance and everyday questions. Work is an agent designed for longer, multi-step work and finished deliverables. Codex remains dedicated to software development and technical work” (help.openai.com/en/articles/20001275, “Updated: 4 hours ago” when checked). Notice the word choice: OpenAI calls Work an agent. Not Chat.
If you're looking for the “ChatGPT agent” or “agent mode” you read about in 2025, it's gone. The help center now says: “ChatGPT agent is no longer available. Use ChatGPT Work for longer, multi-step tasks and finished deliverables” (checked 2026-09-14, “Updated: 30 days ago”). The original launch post carries its own banner: “This launch post is outdated. For current agentic work and team workflows, explore: Workspace agents / ChatGPT Work.” Operator folded into agent mode first; agent mode folded into Work. Two renames in about thirteen months.
What Work does: Tasks (one-time or recurring jobs you assign and walk away from), event-triggered tasks (start themselves, limits below), a Plan mode that shows its approach before starting, a Sites feature, and a built-in browser on desktop. It draws on connected apps, still commonly called connectors, now gathered into a plugins directory. For Business, Enterprise, Edu, or Teachers plans, there's also Workspace agents: named, reusable agents (Spark, Slate, Tally, Scout) that run on a schedule against tools like Slack, Google Drive, Salesforce, and Notion, under admin permissions. Still a research preview.
None of this surprised attentive users. On July 17, 2025, the day the launch post went up, Hacker News commenter fsndz called it early: “By that definition, the ChatGPT app is now an AI agent... it is something new: it is agentic” (fsndz, July 17, 2025). An earlier version of the product, but the framing held up.
The Documented Ceilings (the Table That Answers the Real Question)
Here's where the decision gets made. Every number below is from OpenAI's own help center, checked 2026-09-14 (last updated about 2026-08-28; OpenAI edits this often). Treat this as a snapshot, not a promise, and recheck it before you plan around it.
| Ceiling | Free / Go | Plus | Business / Edu | Pro / Enterprise |
|---|---|---|---|---|
| Active Scheduled Task limit | 3 | 5 | 10 | 15 |
| Recurring task frequency | Once per day, flexible window only, no exact time | Up to once per hour, exact delivery times | Up to once per hour, exact delivery times | Up to once per hour, exact delivery times |
| Event-triggered tasks available? | No | Yes | Yes (not in FedRAMP workspaces) | Yes |
Plain consequences:
- Only three apps can start an event-triggered task: Gmail, Slack, GitHub. Trigger lives elsewhere, ChatGPT can't start the job.
- Combined, event-triggered tasks max out at 30 runs an hour and 720 a day, across all of them, not per task.
- Need a sixth always-on automation on Plus? You're already out of slots.
- An action that sends a message or changes external data can pause for approval before finishing. Documented, not a bug, but “fully unattended” has a ceiling too.
That is the combined ceiling across all of your event-triggered tasks, not a per-task allowance. The hourly ceiling is 30, also combined.
The closest first-party answer to “how much can it do unwatched” is the default permission tier: auto-read, but ask before a message, a deletion, a purchase, or a security change. Admins can adjust that per app.
One wrinkle: OpenAI's own documentation isn't fully consistent on the Business/Edu-versus-Pro/Enterprise split, covered in why ChatGPT Tasks stop sending notifications. Check your account's Scheduled Tasks page for your exact cap.
You Do NOT Need a Dedicated Agent Platform Yet If...
- You're running fewer recurring automations than your plan's active-task ceiling (3 Free/Go, 5 Plus, 10 Business/Edu, 15 Pro/Enterprise). Under that number, ChatGPT's own Tasks already cover you.
- Every trigger you need lives in Gmail, Slack, or GitHub, the entire list of apps that can start an event-triggered task today.
- You don't need custom branching logic. ChatGPT has no user-built if/then workflow branch documented anywhere; its “logic” is the model reasoning inside a Task.
- It's one person's workflow, not a shared team one. Workspace agents, the admin-governed layer, is Business/Enterprise/Edu/Teachers only and still a research preview.
When ChatGPT's Ceilings Mean You Do Need Zapier, Make, n8n, or Lindy
- You've hit your plan's active-task ceiling and would need to retire an automation to make room for a new one.
- Your trigger app isn't Gmail, Slack, or GitHub (table below has the integration numbers).
- You need real conditional branching, which ChatGPT documents none of.
- You need error handling that doesn't rely on your own approval click.
- You need a shared, team-owned workflow outside a research preview.
One thing worth flagging: Zapier, Make, and n8n all give you a countable unit to budget against, a task, a credit, an execution. ChatGPT gives you none of that. Work draws on the same consumption allowance as Codex, and OpenAI doesn't publish a dollar price for it, which makes it genuinely hard to compare on cost.
| Dimension | ChatGPT Work / Workspace agents | Zapier | Make | n8n | Lindy |
|---|---|---|---|---|---|
| Integration breadth (vendor-stated) | 1,400+ plugins overall; only 3 apps trigger a task | 9,000+ apps | 3,000+ apps (also 1,000+ elsewhere on the same page) | not stated on the vendor's page | 1,000+ integrations, plus MCP |
| Entry paid price | not stated on the vendor's page | $29.99/mo, or $19.99/mo annual, 750 tasks/mo | $12/mo at 10,000 credits/mo (billing cadence not stated) | 20 EUR/mo annual, 2,500 executions/mo | $29.99/mo per user, 3,000 credits |
| Branching / conditional logic | not documented as a feature | Yes, Filters & Paths (paid) | Yes, Routers & filters (all plans) | unlimited steps per execution | not stated on the vendor's page |
| Error handling / retries | approval pause only | not stated on the vendor's page | named error-handler modules, credit-exempt | automatic retries listed, tier unspecified | not stated on the vendor's page |
| Team / shared workflows | Workspace agents, Business/Enterprise/Edu/Teachers only, preview | unlimited users, paid tiers | unlimited users, paid tiers | shared projects, every tier | shared credit pool |
From here, the next step is choosing between Zapier, Make, and n8n. Or work out how to evaluate a dedicated agent tool.
Where This Gets Confusing (the Honest Caveats)
Some of what you'll read elsewhere is out of date, not because the writers were careless, but because OpenAI changes this feature set faster than most blogs get republished. A post arguing ChatGPT “isn't a real agent,” written before this year's renames, is now stale.
The terminology alone has moved twice: Connectors folded into a unified plugins directory in July 2026, and agent mode, the mid-2025 flagship, was retired about thirteen months later. Check the date on anything you read about this, including this article. If you'd rather have someone else recheck it for you when OpenAI changes it again, that's what the AgentsExplained newsletter does, one dated verdict at a time.
Worth keeping in view: a user on X described ChatGPT Tasks in August 2025 as “kinda alive, but broken and hidden” (@anatolt, 2025-08-20). That's a reliability complaint about Tasks, not the agent question, but it's a fair reminder that a documented ceiling and day-to-day reliability are two different things.
FAQ
Does ChatGPT count as an AI agent?Yes. OpenAI's own help center states that “Work is an agent designed for longer, multi-step work and finished deliverables” (checked 2026-09-14). The vendor already answered this in its own documentation.
What is an example of an AI agent?A Scheduled Task in ChatGPT Work: set a goal and a trigger, and it runs without re-prompting. On the team side, OpenAI's named Workspace agents (Spark, Slate, Tally, Scout) run on a schedule against company tools.
Is ChatGPT a model or agent? Neither exactly, different layers. ChatGPT is the product; the underlying models power it; Work is the agentic mode inside it, which OpenAI itself labels an agent (help.openai.com/en/articles/20001275).
Is ChatGPT a bot or AI?It's an AI product, not a simple rules-based bot. No first-party page phrases it this way, so we'll stay with what's documented: an AI product with a genuinely agentic mode built in.
The Bottom Line
So, is ChatGPT an AI agent? Yes, by OpenAI's own words. Is it enough of an agent for your job? That depends on whether you fit inside its documented ceilings: three to fifteen active tasks depending on plan, three trigger apps, no branching, solo use unless you're on a research-preview business tier. Check yourself against that list before paying for anything else.
For one honest breakdown like this a week, including the “skip this tool” verdicts, that's roughly what the AgentsExplained newsletter covers.