ChatGPT Workspace Agents
ChatGPT Workspace Agents is OpenAI's product for creating reusable agents for repeatable organizational work. Builders configure instructions, models, files, skills, apps, custom MCP tools and optional memory, then publish access privately, by link, to groups or through a workspace directory. Published agents can run in ChatGPT and, when configured, through Slack, schedules or API triggers.
Origin and context
OpenAI launched workspace agents on 22 April 2026 as a research preview and described them as an evolution of GPTs powered by Codex cloud execution. On 22 May its release notes announced general availability for ChatGPT Business, Enterprise and Edu. Later Slack documentation named OpenAI among the launch platforms for Add to Slack, confirming a separately operated deployment channel rather than only an OpenAI demonstration.
Why it matters
The product turns an agent definition into a shared, versioned organizational resource instead of leaving each user to recreate a prompt and integrations. The same workflow can combine approved context, tools and write controls, be maintained by teammates, and be invoked from several channels. That convenience also centralizes consequential choices about who may publish, whose credentials are used, which actions require approval and how unattended runs are monitored.
Example
A procurement team could publish an agent that receives a vendor request, reads an approved policy file, queries sanctioned data sources, drafts a risk memo and opens a review ticket. Colleagues might invoke it in ChatGPT, while a scheduled run checks outstanding requests. The team should use a scoped service account, constrain connector actions and keep approval enabled for edits or messages; publishing the workflow does not validate its conclusions.
How it differs
Cloud Agents
Cloud agents are the broader pattern of running agent work remotely. ChatGPT Workspace Agents is one branded product whose defining scope also includes reusable publication, organizational sharing, channels, schedules and administration.
Agent sandboxes
An agent sandbox is an isolated execution environment for files, processes, tools or network access. It is one possible runtime layer; it does not by itself provide the product's directory, sharing policy, connected identities, schedules or analytics.
Claude Cowork
Claude Cowork is Anthropic's user-facing surface for handing off multi-step knowledge work across selected files and tools. It can continue in the cloud and serve teams, but it is not OpenAI's separately named, published Workspace Agent object.
Agentic Workflows
An agentic workflow is a vendor-neutral way to organize model decisions and tool actions. A ChatGPT Workspace Agent can implement such a workflow, but the product name should not replace the general pattern.
Maturity and evidence
Maturity is 3. The product has a dated launch, a subsequent GA announcement for three managed-workspace plans, maintained operational and API documentation, an independent Slack deployment relationship, press recognition and external security analysis. It is not rated 4 because it is only months old, product surfaces are changing, official pages retain conflicting preview language, and the reviewed productivity and customer-result claims come from OpenAI or quoted early users rather than neutral comparative studies.
Limits and open questions
Access controls and approvals reduce risk but do not prove safe execution. Agent-owned connections can expose a shared credential's data and actions to everyone allowed to invoke the agent; connector constraints govern requested actions, not all returned content. Scheduled, Slack and API-triggered runs may act without an operator watching each step. In June 2026 OpenAI fixed the reported AgentForger builder CSRF after a lab proof of concept, underscoring the need to audit the whole agent lifecycle. Availability, API response behavior, pricing and supported channels remain changeable. Teams must test outputs and authorization boundaries rather than treat memory, analytics or governance labels as evidence of reliability or compliance.
Related terms
References
- Introducing workspace agents in ChatGPTOpenAI · 2026-04-22 · class A
- ChatGPT Workspace Agents for Enterprise and BusinessOpenAI · 2026 · class A
- ChatGPT Enterprise & Edu - Release NotesOpenAI · 2026-05-22 · class A
- Trigger workspace agent runsOpenAI · 2026 · class A
- Four things you need to know about OpenAI's new workspace agents for ChatGPTIT Pro · 2026-04-24 · class B
- Anyone can now build Agents on Slack: Introducing Add to SlackSlack · 2026-08-20 · class A
- OpenAI Fixes ChatGPT Agent Flaw That Could Let Attackers Forge an AI InsiderSecurityWeek · 2026-07-23 · class B
- AgentForger: A Single Link That Forged a Rogue ChatGPT AgentCloud Security Alliance · 2026-07-24 · class B
- Claude CoworkAnthropic · 2026 · class A
- The next evolution of the Agents SDKOpenAI · 2026-04-15 · class A
Last updated: 2026-09-07
This term is also covered in the Skills Atlas as ai agent design skill.
This term is also covered in the Skills Atlas as workflow orchestration skill.
This term is also covered in the Skills Atlas as agent state management skill.
This term is also covered in the Skills Atlas as agent sandboxing skill.
This term is also covered in the Skills Atlas as low code ai automation skill.