OpenAI hands US teams ChatGPT agents that work on their own

OpenAI started letting eligible US organizations launch business agents on ChatGPT, a feature it calls workspace agents. Announced April 22, 2026, the agents run on OpenAI's Codex model and respond to questions on an organization's behalf. They're built for teams on the Business, Enterprise, Edu and Teachers plans, with the goal of automating workflows that span several steps.

The agents connect to third-party software including Slack, Salesforce and Microsoft tools. They can also run without a human in the loop — on a schedule or triggered by events such as API calls — while staying inside the permissions an organization sets. OpenAI points to lead qualification and report generation as early use cases.

Administrators get role-based access controls and monitoring, so they decide who builds what and can watch what the agents do. Constraints on agent actions and approval checkpoints let a person sign off before an agent does something consequential.

Pricing starts with a free window: workspace agents arrived as a research preview and stay free until May 6, 2026. After that, the feature shifts to credit-based pricing, meaning organizations pay for how much work their agents perform rather than a flat fee.

OpenAI expects workspace agents to gradually displace custom GPTs as the organizational standard. Custom GPTs will keep working for individual users, but OpenAI plans to phase them out for enterprise customers and push businesses toward the agent format. The company first showed ChatGPT could take actions — not just write text — with a general-purpose agent for individual users in July 2025.