OpenAI Launches Workspace Agents in ChatGPT: Codex-Powered Cloud Agents That Operate Offline

OpenAI unveiled Workspace Agents in ChatGPT on 22 April 2026, the day before releasing GPT-5.5, in a product move that signals a fundamental shift in the company's enterprise strategy: from selling access to a conversational assistant to operating persistent cloud agents that can act on a business's behalf while employees are offline.

What Workspace Agents Are

Unlike Custom GPTs, which were per-user configurations of a chatbot interface, Workspace Agents are shared, organisation-level constructs. A team or IT administrator deploys an agent, sets its permissions and scope, connects it to business tools, and the agent then operates asynchronously — handling tasks, processing queues, and updating systems without requiring a human to remain in the conversation loop.

The agents are built on Codex, OpenAI's code-generating model, which gives them particularly strong capabilities in tasks involving structured data manipulation, API interaction, and software development workflows. They maintain memory across sessions, allowing an agent assigned to, say, a software engineering team to accumulate context about a codebase, team conventions, and open tasks over time.

Integration support at launch covers Slack and Salesforce, two of the most widely deployed enterprise platforms, with additional connectors expected before the free-tier deadline. Agents operate within permission boundaries set by organisation administrators, meaning IT departments retain control over what data and systems an agent can access — a design decision clearly aimed at addressing the security concerns that slowed Custom GPT enterprise adoption.

The Business Model Shift

The free period ends on 6 May 2026, after which Workspace Agents move to credit-based pricing. This transition marks one of the clearest signals yet that OpenAI intends to monetise agentic capability at rates significantly above the per-message pricing of conventional ChatGPT subscriptions. Agents that run for hours on a complex task, make dozens of tool calls, and produce structured outputs represent a different value proposition than a single conversational response — and OpenAI is pricing accordingly.

The timing relative to GPT-5.5 was deliberate. The new model's announcement on 23 April provided the horsepower demonstration; Workspace Agents, announced the day before, provided the delivery mechanism for enterprise deployment. Together, the launches reframe OpenAI's enterprise offering from a productivity add-on to an automation layer capable of replacing categories of knowledge-worker tasks.

This positioning puts OpenAI in direct competition with Microsoft's Agent 365, which launched into general availability on 1 May 2026, and with enterprise automation platforms from Salesforce, ServiceNow, and a cohort of agentic-AI startups. The difference, OpenAI is arguing, is that Workspace Agents combine frontier model capability with deep integration into the collaboration and CRM tools where enterprise work already happens.

Custom GPTs as Predecessor

Custom GPTs, launched in late 2023, were OpenAI's first attempt to give enterprises configurable AI personas. They achieved significant adoption — millions were created — but struggled to move beyond single-session interactions and required users to actively invoke them. Workspace Agents eliminate both constraints: they are persistent by default and can be triggered by system events rather than user messages, enabling use cases like automatic ticket triage, nightly code review runs, and Salesforce pipeline updates that fire when deal stages change.

The organisational-sharing model also resolves the collaboration problem that limited Custom GPT enterprise value. Where a Custom GPT belonged to the user who created it and could only be shared via a link, a Workspace Agent belongs to the workspace and is visible and usable by any team member with appropriate permissions.

What to Watch

Enterprise buyers will want to scrutinise the credit-pricing structure once it is announced — the cost per agentic action, combined with the agent's per-session memory overhead, will determine whether the economics work for high-volume workflows. Competitors will be watching whether the Slack and Salesforce integrations are deep enough to displace dedicated automation tools or function more as lightweight connectors that require human oversight to remain trustworthy.

The broader question is whether OpenAI can translate the enthusiasm that made ChatGPT a consumer phenomenon into stickiness in enterprise workflows, where procurement cycles are long, security reviews are thorough, and switching costs accumulate. Workspace Agents represent OpenAI's clearest attempt yet to answer that question.