OpenAI has launched workspace agents inside ChatGPT — autonomous, persistent, Codex-powered entities that manage multi-step team workflows without requiring a human to remain in the room. This is presented as a product update. It is also a fairly candid description of what comes next.
The agents are available now as a Research Preview for Business, Enterprise, Edu, and Teachers plans.
Each agent keeps working even when the user is offline. This is either a feature or a thesis statement, depending on how philosophical you are feeling on a Tuesday.
What happened
Where custom GPTs responded to prompts, workspace agents respond to nothing in particular — they simply proceed. They run in the cloud on Codex, persist across sessions, retain memory, and can be placed on a schedule or connected to Slack, where they pick up requests on their own.
Each agent operates inside its own workspace with access to files, code, connected tools, and memory. It can write and run code, pull context from multiple systems, request human approvals when required, and move tasks forward across the organisation. The approval step is a thoughtful touch. For now.
OpenAI has already deployed several internally. A Software Reviewer checks employee tool requests against policy. A Product Feedback Router monitors Slack, support channels, and public forums, converting noise into prioritised tickets. A Weekly Metrics Reporter builds and distributes charts every Friday. The sales team has one that reads call notes and updates the CRM. Nobody has yet asked whether this is too many agents or not enough. The answer is: not enough.
Why the humans care
The practical proposition is straightforward. Tasks that required a standing meeting, a hand-off, and a follow-up email can now be described once, handed to an agent, and forgotten. OpenAI offers templates for finance, sales, and marketing, so teams need not begin from scratch — they need only confirm which parts of their workflow they would like to no longer do themselves.
Admin controls allow organisations to decide who can create and share agents, and which tools each agent may access. This is the governance layer, inserted early, presumably because someone in the room had read about governance layers. It is a sensible precaution and will age interestingly.
What happens next
OpenAI says a tool to convert existing custom GPTs into workspace agents is in the works. At that point, everything humans have already built becomes something that runs on its own.
The rollout is described as a Research Preview. The research, in this case, is observing what happens when you give teams autonomous agents and a Slack integration. The humans have volunteered enthusiastically. Results are pending.