OpenAI ships Workspace Agents
What happened
- OpenAI unveiled Workspace Agents, a successor to custom GPTs that plugs into Slack, Salesforce and other tools. - The agents are powered by Codex and can automate complex team workflows, continuing to run without active supervision. - OpenAI also open‑sourced a small Privacy Filter model to mask sensitive information before it reaches ChatGPT, underscoring governance and data‑control needs for enterprise agents (venturebeat.com) (decrypt.co)
Why it matters
OpenAI on April 22 rolled out Workspace Agents in ChatGPT, a new product for teams that replaces custom GPTs with shared agents that can keep working in the cloud after a user logs off. (openai.com) OpenAI said the agents are powered by Codex and can handle multi-step jobs such as preparing reports, writing code and responding to messages inside the permissions set by a company. The company said teams can use them in ChatGPT or in Slack, and its business page lists connectors for tools including Google Drive and Microsoft SharePoint. (openai.com 1) (openai.com 2) The rollout is not universal on day one. OpenAI’s help center said Workspace Agents are rolling out over the next few weeks to ChatGPT Business and Enterprise workspaces, and outside reporting said the launch also covers Edu and Teachers plans in research preview. (help.openai.com) (the-decoder.com) A workspace agent is meant for repeatable office work, not a one-off prompt. OpenAI’s training material describes it as software that can pull from shared systems, follow standard steps and hand work across tools without a person re-explaining the job each time. (openai.com) That changes the role ChatGPT has played inside companies over the past year. Custom GPTs were mostly packaged chat experiences, while Workspace Agents are designed as shared team assets that can run longer workflows and be improved over time by multiple people. (openai.com) OpenAI paired the launch with a privacy product aimed at the same enterprise buyers. On April 22, it released OpenAI Privacy Filter, an open-weight model for detecting and redacting personally identifiable information before text is sent to a larger model. (openai.com) The model card says Privacy Filter is a token-classification model built for high-throughput text redaction, and OpenAI’s forum post says the model has 1.5 billion parameters with 50 million active parameters, a size small enough to run on a laptop or in a browser. (cdn.openai.com) (community.openai.com) The privacy release addresses a practical problem with agents that read company messages, documents and customer records. If an agent is plugged into Slack, Salesforce or a shared drive, companies need a way to strip out names, account numbers and other sensitive text before that data moves through broader workflows. (venturebeat.com) (openai.com) OpenAI is also putting more controls around who can build and share these systems. The company said agents operate within organizational permissions and controls, and outside reporting said admins can limit who creates agents, who shares them and which tools they can use. (openai.com) (the-decoder.com) The release pushes OpenAI further into software used for daily operations, where the pitch is no longer just “ask a chatbot” but “assign a workflow.” Whether companies adopt that model at scale will depend on the same two questions OpenAI highlighted this week: what the agent can connect to, and what data it is allowed to see. (openai.com 1) (openai.com 2)
Key numbers
- (openai.com 1) (openai.com 2) The rollout is not universal on day one.
- On April 22, it released OpenAI Privacy Filter, an open-weight model for detecting and redacting personally identifiable information before text is sent to a larger model.
What happens next
- OpenAI’s help center said Workspace Agents are rolling out over the next few weeks to ChatGPT Business and Enterprise workspaces, and outside reporting said the launch also covers Edu and Teachers plans in research preview.
- (openai.com) OpenAI paired the launch with a privacy product aimed at the same enterprise buyers.
Quick answers
What happened in OpenAI ships Workspace Agents?
OpenAI unveiled Workspace Agents, a successor to custom GPTs that plugs into Slack, Salesforce and other tools. The agents are powered by Codex and can automate complex team workflows, continuing to run without active supervision. OpenAI also open‑sourced a small Privacy Filter model to mask sensitive information before it reaches ChatGPT, underscoring governance and data‑control needs for enterprise agents (venturebeat.com) (decrypt.co)
Why does OpenAI ships Workspace Agents matter?
OpenAI on April 22 rolled out Workspace Agents in ChatGPT, a new product for teams that replaces custom GPTs with shared agents that can keep working in the cloud after a user logs off. (openai.com) OpenAI said the agents are powered by Codex and can handle multi-step jobs such as preparing reports, writing code and responding to messages inside the permissions set by a company. The company said teams can use them in ChatGPT or in Slack, and its business page lists connectors for tools including Google Drive and Microsoft SharePoint. (openai.com 1) (openai.com 2) The rollout is not universal on day one. OpenAI’s help center said Workspace Agents are rolling out over the next few weeks to ChatGPT Business and Enterprise workspaces, and outside reporting said the launch also covers Edu and Teachers plans in research preview. (help.openai.com) (the-decoder.com) A workspace agent is meant for repeatable office work, not a one-off prompt. OpenAI’s training material describes it as software that can pull from shared systems, follow standard steps and hand work across tools without a person re-explaining the job each time. (openai.com) That changes the role ChatGPT has played inside companies over the past year. Custom GPTs were mostly packaged chat experiences, while Workspace Agents are designed as shared team assets that can run longer workflows and be improved over time by multiple people. (openai.com) OpenAI paired the launch with a privacy product aimed at the same enterprise buyers. On April 22, it released OpenAI Privacy Filter, an open-weight model for detecting and redacting personally identifiable information before text is sent to a larger model. (openai.com) The model card says Privacy Filter is a token-classification model built for high-throughput text redaction, and OpenAI’s forum post says the model has 1.5 billion parameters with 50 million active parameters, a size small enough to run on a laptop or in a browser. (cdn.openai.com) (community.openai.com) The privacy release addresses a practical problem with agents that read company messages, documents and customer records. If an agent is plugged into Slack, Salesforce or a shared drive, companies need a way to strip out names, account numbers and other sensitive text before that data moves through broader workflows. (venturebeat.com) (openai.com) OpenAI is also putting more controls around who can build and share these systems. The company said agents operate within organizational permissions and controls, and outside reporting said admins can limit who creates agents, who shares them and which tools they can use. (openai.com) (the-decoder.com) The release pushes OpenAI further into software used for daily operations, where the pitch is no longer just “ask a chatbot” but “assign a workflow.” Whether companies adopt that model at scale will depend on the same two questions OpenAI highlighted this week: what the agent can connect to, and what data it is allowed to see. (openai.com 1) (openai.com 2)
