Workspace Agents.

OpenAI shipped workspace agents this week. GPTs that do the work while you sleep.

I’ve been running my own version of this for months. Claude Code on a cron schedule. Agents that process my vault at 6pm, 6:30pm, and 7:30pm. No prompts. No chat window. Work gets done in the background.

So when I read the announcement, the feature didn’t surprise me. What surprised me was that OpenAI finally shipped it as a product instead of a demo.

What actually changed

The headline is “workspace agents in ChatGPT.” It is three things at once:

1. Agents that run in the cloud on their own schedule
2. Approval gates for anything sensitive like sending emails or touching customer data
3. Persistent context across tools, Slack, files, and systems

Chat was always the wrong primitive for team work. You don’t want to type into a box every time a report needs to go out. You want the report to write itself on Friday afternoon, land in your inbox, and flag the things that need a human.

That’s what workspace agents do. OpenAI launched with four examples: a software reviewer that checks requests against policy, a product feedback router watching Slack, a weekly metrics reporter, and a lead outreach agent that researches prospects.

All of them run in the background. None of them need you online.

Why the approval gate matters more than the agent

Everyone will focus on “agents can do work for you.” That’s not the interesting bit.

The interesting bit is that admins can require approvals for sensitive actions. Sending an email? Approval required. Touching a customer record? Approval required. Without that, this is a toy. With it, a team can actually run it in production.

Every AI tool I’ve seen fail inside a company fails for one of two reasons. Too much access and legal blocks it. Too little access and it does nothing useful. Approvals sit in the middle. The agent drafts. A human signs off on what matters.

I’ve been running that exact pattern on Claude Code. Agents draft, humans approve. It works.

How to actually use this

If you are on ChatGPT Business, Enterprise, Edu, or Teachers: workspace agents are free until May 6, 2026. After that it’s credit-based.

Don’t build 10 agents. Build one. Pick the most annoying repeatable thing your team does every week. The one three people touch, nobody enjoys, and it always slips to Friday afternoon.

Build that. Wire the approval gate for anything that goes external. Run it for two weeks and watch what breaks. Fix it. Then build the next one.

That’s how I built my vault system. I didn’t design it up front. I let the friction tell me what to automate next.

Where this falls short

Codex powers all of it. You are betting on OpenAI’s model layer for every agent you ship. If you already run on Claude or Gemini for production workflows, this is another silo to manage.

“Research preview” means the feature will be half-broken for six months. Expect permission bugs, approval flow bugs, and quiet failures where the agent stops running and nobody notices.

Credit-based pricing from May 6 will punish teams that over-built during the free window. Budget for it now.

Where this lands

Chat is not how work happens. Work happens on a schedule, with context, with someone signing off on the bits that matter.

If you are still prompting ChatGPT by hand for anything repeatable, you are doing it wrong.