What Can ChatGPT Agents Actually Do in 2026?

Agent Mode Is Gone — Here's What Replaced It

OpenAI's own help center is direct about it: “ChatGPT agent is no longer available. Use ChatGPT Work for longer, multi-step tasks and finished deliverables.” The retirement notice appeared in early August 2026, pointing users toward Work for general tasks and toward a separate cloud browser tool for pure browser-driven workflows. If you'd bookmarked instructions for typing /agent into the composer and it stopped doing anything, this is why.

OpenAI launched ChatGPT Work on July 9, 2026, the same day as the GPT-5.6 model family that powers it. It isn't simply Agent Mode with a new label — OpenAI folded agent-style task execution into a broader mode that sits alongside everyday Chat and the developer-focused Codex, giving ChatGPT three distinct modes instead of one chat window with a hidden agent toggle buried in a menu. Notably, OpenAI's own documentation doesn't explicitly describe it as a straight rename, so it's more accurate to think of Work as Agent Mode's successor with an expanded scope than as the exact same feature under a new name.

Switching From Chat to Work

A simple switcher at the top of the composer moves you from Chat to Work — no separate menu to dig through.

On web and mobile, there's a Chat/Work switcher directly at the top of the message composer — select Work and you're in, no settings page required. On desktop, OpenAI merged the separate Codex app and the main ChatGPT app into one new application at the July 9 relaunch, renaming the previous standalone app “ChatGPT Classic.” If you're still using Classic, Work won't be there at all; you'll need the new unified desktop app to access it.

Work isn't available on the Free or Go plans in the old Agent Mode sense, though OpenAI's pricing page shows limited Work access for Free and Go accounts specifically inside the new desktop app. Plus, Pro, Business, and Enterprise get full access, drawing from the same agent-usage credit system that Codex already used.

Real Deliverables, Not Just Chat Replies

Work is built for tasks that end in a finished file rather than a conversation: a spreadsheet, a slide deck, a written report, or even a small working web app. It stays with a multi-step project for as long as it needs — OpenAI describes it as able to remain on complex work for hours — pulling in your connected apps and uploaded files as it goes rather than treating each step as an isolated question.

It connects outward through a Plugins directory (an expanded, renamed version of the old app directory) covering Slack, Microsoft Teams, Google Drive, Gmail, SharePoint, Salesforce, GitHub, Linear, Figma, Adobe Acrobat, Zoom, Canva, Dropbox, and dozens more, primarily on Business and Enterprise plans. It can also run tasks on a schedule — a recurring weekly report, for instance — rather than only responding to something you type in the moment.

What a Real Work Task Looks Like

Say you need a competitive pricing snapshot before a Monday planning meeting. Switch to Work and describe the outcome, not the steps: “Research the pricing pages of these five competitors, note any changes from what's in this spreadsheet I'm attaching, and build an updated comparison sheet.” Work plans out the sub-tasks, visits each site, cross-references against your uploaded file, and returns a finished spreadsheet rather than a wall of text you'd have to organize yourself.

If something in the task requires a judgment call — a competitor's pricing page is ambiguous, or a figure doesn't match your existing data — it pauses and asks rather than guessing silently and handing back something wrong. That confirmation step is deliberate: OpenAI built Work to interrupt for anything consequential or unclear rather than optimizing purely for speed.

What Got Lost in the Transition

The old Agent Mode had one capability that hasn't cleanly carried over: driving a logged-in browser session and handing control back to you mid-task. Agent Mode used to spawn a real browser, visually interpret the page, click through interfaces, and pause at logins or payments so you could finish that one step by hand before it continued — which is what made workflows behind a sign-in wall (product configurators, account settings, checkout flows) possible at all.

Work covers research, public browsing, files, and document creation well, but according to developers flagging the gap on OpenAI's own forum in early August 2026, it doesn't reproduce that specific handoff — driving a logged-in session and pausing for you to complete a sensitive step before resuming. OpenAI's current guidance for that specific need points to a separate cloud browser tool inside ChatGPT rather than Work itself. The forum discussion also noted there was no advance notice or migration guide for the change, so if a workflow you relied on depended on that logged-in handoff, it's worth testing the cloud browser directly rather than assuming Work covers it the same way.

It Still Checks In Before Anything Consequential

The safeguards that defined Agent Mode carried over into Work. It pauses and asks for confirmation before high-impact actions — submitting a form, making a purchase, sending a message on your behalf — and you can watch its progress in real time or interrupt and redirect it mid-task. OpenAI's documentation is explicit that these measures reduce risk without eliminating it entirely, so it's still worth supervising a task rather than assuming a confirmation prompt will always catch every problem, particularly on unfamiliar sites where a hidden instruction embedded in a webpage could attempt to steer the agent off-task. That risk isn't hypothetical — OpenAI's own documentation for the previous Agent Mode described exactly this scenario: a malicious comment on a page instructing the agent to retrieve a password reset code and send it somewhere it shouldn't, which is precisely the kind of action the confirmation step is designed to catch before it happens.

Frequently Asked Questions