OpenAI released a code change in the public repository of the coding agent Codex on August 26, 2026, which prepares a new persistent mode. The mode allows the agent to continue working without a fixed time limit, whereas previous settings would stop after minutes or hours. As the magazine Wired reports, the agent will be able to create its own follow-up tasks in the future and continue them over multiple sessions.
Pull request reveals technical details of the mode
The merged pull request in the source code repository of Codex adds “persistent” as a new value to the reasoning effort protocol and the TypeScript SDK types. In the terminal interface, the setting appears as “Persistent”; internally, Codex translates the value for the responses API to “disabled.” According to the commit descriptions, the testing includes parsing and serializing the setting, translating requests to remote models, and selection in the command line interface.
More interesting than the technical implementation, according to Wired, is the internal instruction in the code: the agent is simply supposed to “keep working until it is put to sleep.” This includes a function referred to as “proactivity,” which allows the agent to independently create follow-up tasks. It is supposed to draw on previous interactions and its knowledge about the user to decide what to work on next and, in some cases, reach out to the user on its own. Already in July, OpenAI demonstrated with the encryption of messages between Codex agents how much communication between multiple Codex instances is now automated.
OpenAI confirms testing but gives no timeline
A spokesperson for OpenAI confirmed to Wired that the company is internally testing the persistent mode but did not announce any imminent broad release. According to the report, the company also acknowledges that permanently active AI systems inherently carry higher risks than agents with a fixed time limit. The source code therefore explicitly limits what the agent is additionally allowed to do in persistent mode: its powers remain unchanged according to the instructions, and interventions outside its own working environment still require user approval.
This move is part of a larger product offensive surrounding Codex. Just in July, OpenAI launched Codex Micro, a dedicated controller for Codex agents, complete with a joystick and reasoning dial. Persistent mode fits into this picture: OpenAI is increasingly turning Codex from a pure chat tool into an independently working colleague that continues tasks even without active supervision.
What will matter is whether OpenAI actually enables continuous operation for paying customers or classifies it as an internal experiment. So far, persistent only appears in the public source code, not in the official product documentation. For developers using Codex in their daily work, this decides whether a coding agent will keep working independently at night and over weekends in the future — or whether OpenAI shelves the feature again after the first public reactions.
