OpenAI has introduced workspace agents in ChatGPT. These are cloud-based assistants powered by the Codex model. They are designed to handle multi-step workflows for teams and connect directly with tools like Slack, Google Drive, Salesforce, and Microsoft apps.
The new system shifts ChatGPT beyond simple prompt responses. Instead of replying to one request at a time, these agents can manage longer tasks that continue running in the background. OpenAI describes this as a step forward from custom GPTs toward full workflow execution for teams.
The feature is currently available as a research preview. It is limited to users on ChatGPT Business, Enterprise, Edu, and Teacher plans.
How Workspace Agents Work
Workspace agents run inside a dedicated environment. They have access to files, connected tools, code execution, and stored memory. This memory carries across sessions, so they can remember context over time.
Unlike earlier custom GPTs that respond only when prompted, workspace agents can plan and carry out multi-step actions. They gather information from different systems, follow defined team processes, and move tasks forward without repeated instructions.
They can also request approval when needed. After completing part of a workflow, they can pause and wait for human confirmation before continuing. This creates a balance between automation and control.
Integration With Business Tools
Agents communicate directly with workplace platforms. Teams can utilize them within ChatGPT or deploy them inside Slack channels. After it is added to Slack, they observe the conversations and reply on their own.
They also support scheduled tasks. An agent can be set to run at particular times or triggered by events. It enables the automation of repetitive workflows without manual intervention in tracking.
A company may schedule weekly reports, data summaries, or routine checks, for example. The agent works in the background and returns results as needed.
Real Use Cases Inside OpenAI
OpenAI has already tested workspace agents internally. The company shared several examples in use across its teams.
A Software Reviewer agent checks employee requests for software tools. It compares them against approved lists and company policies. If something is missing or non-compliant, it creates IT tickets automatically.
An accounting-focused agent supports month-end closing tasks. It prepares journal entries, helps with balance sheet reconciliations, and reviews financial variances. This reduces manual work during reporting cycles.
A Product Feedback Router scans Slack channels and support forums. It collects user feedback, organizes it, and turns it into structured weekly action items for product teams. This helps teams respond faster to user input without manually sorting messages.
Control, Security, and Safeguards
OpenAI has added safeguards to manage risk. Workspace agents include protection against prompt injection attacks. They also support monitoring through a Compliance API, which gives visibility into agent activity.
Administrators can set limits on what data each agent can access. They can also control which tools the agent is allowed to use. Sensitive actions like sending emails or creating calendar events can require human approval before execution.
This structure is meant to reduce errors and prevent unwanted actions while still allowing automation.
Availability and Pricing
Workspace agents are currently free during the research preview phase. This period runs until May 6, 2026.
After that, OpenAI plans to introduce credit-based pricing for business and enterprise users. The final cost structure has not been fully detailed yet.
Existing custom GPTs will remain available during the preview period. Users can continue using them alongside the new workspace agents.
What This Means for Teams
Workspace agents shift ChatGPT closer to a full work automation system. Instead of acting as a chat assistant, it now handles ongoing tasks across multiple tools.
For teams, this changes how routine work is managed. Tasks that once needed direct human attention can now run in the background with little input. Instead of asking questions, the focus shifts to workflows that execute.
