OpenAI's bet on shared agents is the quietest shift in enterprise AI this year

On April 22, 2026, OpenAI announced a product category that looks, on its face, like a polished upgrade to the GPTs it introduced in late 2023. But these workspace agents are something more structural: an attempt to move enterprise AI from a personal productivity tool into a persistent, shared organizational layer.

The agents run on Codex, OpenAI's model tuned for code generation and task execution, and live in the cloud instead of on a user's device. They can be scheduled to run on their own, hook into tools like Slack and CRMs, and remember what happened in past sessions. The key point: they are built to be shared across a team. One person builds an agent, publishes it to the company agent library, and everyone else can use, customize, and improve it over time.

From GPTs to agents: what changed

The original GPTs were lightweight, prompt-based helpers tied to individual ChatGPT conversations. No persistent state. No cross-session memory. No way to run on a schedule. The workspace agents scrap all those limits: they can write and execute code, talk to external APIs, remember past interactions, and fire off on a timer. OpenAI says they 'are powered by Codex in the cloud, which gives them access to a workspace for files, code, tools, and memory.'

This is a fundamentally different architecture, one that shifts the agent from a conversational interface to an autonomous background worker. The company says agents 'continue to work even when you are away' and can be set to 'run on a schedule or deployed in Slack to handle requests as they arrive.'

Real workflows, real teams

OpenAI showed the product with concrete internal use cases. The sales team uses an agent that 'gathers information from call notes and account research, qualifies new leads, and writes follow-up emails directly in a salesperson's inbox.' The accounting team built an agent that 'prepares key items of the monthly close, from journal entries to balance sheet reconciliations to variance analysis.' The product team deployed one that 'proactively answers employee questions in Slack channels.'

These are not toy demos. They represent workflows that used to demand coordination across multiple tools, manual data gathering, and human judgment. Offloading them to a persistent agent lets team members focus on higher-value decisions. Rippling's AI Engineering lead Ankur Bhatt put it: 'What used to take salespeople 5-6 hours per week now runs automatically in the background on every opportunity.'

The governance question

Autonomous agents need guardrails. OpenAI built controls at multiple layers. Agent creators can set which tools and data the agent can access, which actions need human approval, and whether the agent can do sensitive operations like sending emails or modifying spreadsheets without confirmation. In Enterprise and Edu plans, admins can manage which user groups can use, create, and share agents, and can monitor execution analytics, total runs, unique users, activity over time.

The company also added a Compliance API that provides visibility into each agent's configuration, updates, and runs. Administrators can also suspend agents if necessary, OpenAI says. Planned features include a full agent inventory in the admin console, showing usage patterns and connected data sources.

The governance model draws on lessons from earlier autonomous agent experiments, where uncontrolled tool access led to costly mistakes. By baking approval gates and audit trails into the platform, OpenAI addresses the main concern that has kept many enterprises from deploying autonomous AI at scale.

Pricing and availability

Workspace agents are available in research preview for ChatGPT Business, Enterprise, Edu, and Teachers plans. They are free until May 6, 2026, after which OpenAI shifts to credit-based pricing, charging per execution rather than per seat. That aligns incentives: customers pay for actual work done, not for potential access. It also caps OpenAI's exposure to runaway compute costs from deeply inefficient agents.

The free preview suggests OpenAI is still gathering data on real-world usage patterns to calibrate pricing. The company will need to balance accessibility with profitability, a tension that has shaped its enterprise product trajectory since ChatGPT Business launched.

What comes next

OpenAI has promised future additions: new triggers for automatically launching tasks, better dashboards for understanding and optimizing performance, more ways for agents to act across business tools, and support for workspace agents in the Codex application. The company also committed to making it easy to convert existing GPTs into workspace agents, preserving the investments teams have made.

The roadmap signals ambition beyond this launch. If workspace agents succeed, they could redefine how enterprises think about AI: not as a chatbot you talk to, but as a workforce you manage. That shift, from interaction to delegation, is what makes today's announcement more consequential than it first appears.