On April 22, OpenAI released a product called workspace agents, formally announcing that custom GPTs are being replaced.
This is not a feature upgrade; it is a shift in logic.
GPTs answer questions; agents work autonomously
The core of custom GPTs: the user asks a question, and the model answers once.
The core of workspace agents: assign it a task, and it runs on its own — you can shut down and sleep while it continues working in the cloud.
Under the hood, it uses the Codex model. Each agent has its own independent workspace containing files, a code execution environment, tool invocation permissions, and cross-session persistent memory. You don't need to re-explain the context every time; it remembers.
More importantly, the design emphasizes "team sharing" — once an agent is built, it can be shared across the entire organization via ChatGPT or Slack, and can be continuously iterated and improved. This is something custom GPTs never achieved.
OpenAI used itself as a testbed
At launch, they shared several agents running internally, a detail that speaks volumes:
- Software Reviewer: Checks whether new requirements comply with company policy
- Product Feedback Router: Monitors Slack and customer service channels, automatically categorizing and routing feedback
- Weekly Metrics Reporter: Automatically generates a metrics report every Friday without human prompting
- Lead Outreach Agent: Researches and scores potential leads, drafts outreach emails
- Accounting Agent: Handles the monthly closing process
These are not demos; they are actually running. OpenAI is testing on itself before selling to others — a familiar rhythm.
Sensitive operations require human approval
One design detail worth noting: operations that have external impact, such as sending emails or creating calendar events, require manual confirmation by default before the agent can execute them. This is a safeguard mechanism and a response to concerns about excessive AI autonomy.
Meanwhile, which files an agent can access and which tools it can invoke are entirely controlled by the organization's administrator. The system also includes built-in prompt injection protection to prevent external content from hijacking agent behavior.
Pricing and timeline
Currently in research preview, available only to Business, Enterprise, Edu, and Teachers plans.
Free until May 6; credit-based billing thereafter.
A GPT conversion tool is under development, allowing existing users to migrate custom GPTs into workspace agents. During the migration period, existing GPTs will not be taken offline immediately.
Three launches on the same day — an interesting coincidence
On the same day (April 22), Google launched Workspace Intelligence at Cloud Next, embedding Gemini into Gmail, Docs, and Sheets. Microsoft had earlier rolled out Copilot Wave 3, also an enterprise agent architecture.
All three are doing the same thing: embedding AI agents into enterprise workflows, turning them into infrastructure as fundamental as Slack.
The difference lies in the entry point — OpenAI uses Codex as the engine and ChatGPT as the gateway; Google uses Gemini, starting from Gmail; Microsoft uses Copilot, starting from Office.
Who will win? Probably not one company winning everything, but different tools for different scenarios. However, the number of users paying credits starting May 7 will be the first real answer.
Sources: OpenAI launches workspace agents that turn ChatGPT from a chatbot into a team automation platform (The Decoder); OpenAI subscribers get new workspace agents to automate complex tasks across teams (SiliconANGLE); CocoLoop, OpenAI updates ChatGPT with Codex-powered workspace agents for teams (9to5Mac)