OpenAI to run AI agents on dedicated computers

OpenAI announced at DevDay 2026 that its AI agents will run on persistent private compute instances assigned to individual agents. OpenAI noted the machines can run continuously, hold local state and access connected resources such as files, sensors and networks.

Enterprises can choose cloud-hosted private instances or on-premises devices that run an agent runtime. Developers will be able to allocate a machine to an agent through a new agent management API. OpenAI presented tools for provisioning, monitoring and revoking access, and the devices will include options for hardware-backed security and encrypted local storage to meet compliance requirements.

OpenAI also announced updates to its developer tools. New SDKs include scheduling and event hooks, memory primitives were expanded so agents can maintain longer-term records, and a policy framework was introduced to govern agent actions and external access. The company previewed a marketplace for agent extensions that add connectors to enterprise systems, internet-of-things devices and third-party services.

The initial rollout will prioritize enterprise customers and partners through an early-access program; broader developer access is planned later in the year. OpenAI provided limited pricing information and indicated costs will vary by configuration, with more details to be released during the beta period.

OpenAI described the changes as addressing two developer needs: continuous operation for ongoing workflows and clearer boundaries for data control. Dedicated computers allow agents to keep sensitive information local rather than repeatedly sending it to a remote model, and management APIs are meant to let organizations audit agent behavior and revoke permissions.

In a statement, an OpenAI representative wrote, “Agents will run continuously, retain necessary local context and interact directly with devices and systems under customer control.” The representative added the company plans to provide monitoring and compliance features alongside the new hardware options.

The announcements build on OpenAI’s earlier work with models that call external tools and hold memory. Previous products allowed developers to extend models through plugins and custom instructions; the new approach places persistent compute and governance tools at the center of agent deployment.

OpenAI did not give a firm general-availability date for the dedicated-computer option or the agent marketplace. OpenAI plans to post more technical documentation, pricing details and sign-up information to its developer site in the coming weeks.