OpenAI dropped something interesting this week — workspace agents inside ChatGPT. Not the chat-based assistants we’ve been tinkering with, but actual Codex-powered agents that run in the cloud and automate complex workflows.
I’ve been testing them for a few days, and they’re genuinely different from the usual “ask a bot, get a text reply” pattern. These agents can pull data from Google Drive, write to a Jira ticket, run a script, and send a Slack notification — all in one sequence, without me sitting there clicking buttons.
What They Actually Do
The core idea is straightforward: you describe a workflow in natural language, and the agent builds a series of steps using Codex (OpenAI’s code-generating model). The agent runs in OpenAI’s cloud, not on your machine, so it can keep executing even after you close your laptop. That’s a big deal for anyone who’s tired of babysitting scripts.
Out of the box, it connects to common tools — Google Workspace, Microsoft 365, Jira, GitHub, Slack, Notion, Salesforce. You authenticate once, and the agent handles the rest. I set up a workflow that triages incoming support emails: it reads the message, checks a knowledge base, creates a Jira ticket if needed, and posts a summary to Slack. Took about 10 minutes to configure.
The agent surfaces what it’s doing in a side panel, so you can watch the steps execute in real time. If something fails — like a missing permission — it pauses and asks for input, rather than silently crashing. That’s a nice touch.
Security and Sharing
Each workspace agent runs in a sandboxed environment with its own API credentials. OpenAI says no customer data is used for training, and the agents respect the same data retention policies as the rest of ChatGPT Enterprise. You can also set granular permissions per agent — who can view, edit, or trigger it.
Sharing is straightforward: publish an agent to your workspace, and teammates can run it with their own credentials. No need to duplicate setup work. Version history is tracked, so if someone breaks the workflow, you can roll back.
Pricing and Availability
Workspace agents are available today for ChatGPT Enterprise and Team subscribers. Enterprise customers get priority execution slots and longer timeouts (up to 30 minutes per run). Team users get 10-minute max runs, which is fine for most workflows but might choke on heavy data pipelines.
Pricing is baked into existing subscriptions — no extra per-agent fee, which surprised me in a good way. OpenAI confirmed they’re not metering execution time (yet). I expect that to change once they see usage patterns, but for now it’s a fair deal.
Where It Falls Short
I’ll be honest: it’s not magic. The agent struggles with ambiguous instructions — if you say “process the spreadsheet,” it might try to open every sheet instead of just the one you meant. Error messages can be cryptic; I got a “step timed out” with no explanation of which step or why.
Also, it only works inside ChatGPT’s interface. There’s no API to trigger agents programmatically, no webhook support, no scheduled runs. You have to manually start each execution. That limits its usefulness for truly unattended automation.
Competitors like Zapier and Make have had visual workflow builders for years, and they let you schedule runs. OpenAI’s agent is more powerful under the hood (Codex understands context better than a drag-and-drop builder), but it’s less mature as a product.
Should You Use It?
If your team already lives inside ChatGPT Enterprise, this is a no-brainer. It reduces context switching and lets non-technical people automate tasks they’d otherwise hand off to engineering. For power users building complex data pipelines, you’ll still want dedicated tools — but for the 80% of office workflows that involve reading, writing, and notifying, workspace agents work well enough.
I’m curious to see where OpenAI takes this. If they add scheduling, webhooks, and an API, it could eat into Zapier’s lunch. If they don’t, it’ll remain a nice add-on for ChatGPT power users. Either way, it’s the most practical automation tool OpenAI has shipped since GPT-4.
Comments (0)
Login Log in to comment.
Be the first to comment!