OpenAI Develops Persistent Codex Agent Mode
Wired reported on August 27 that OpenAI is developing a "Persistent mode" for its Codex command-line coding agent, based on code appearing in the tool. Gizmodo reported that, if released in its current form, the mode could keep Codex running until a user puts it to sleep, with controls for reasoning effort and an apparent "Proactivity" capability for creating follow-up tasks.
Wired reported on August 27 that code for a "Persistent mode" has begun appearing in the command-line version of OpenAI's Codex coding agent. Gizmodo, citing Wired's reporting, described the feature as an always-on mode that could continue operating until a user actively puts the agent to sleep.
The reported code also appears to include a "reasoning effort" setting, which Gizmodo described as a way for users to choose how much compute the model can use while working through a task. A separate apparent feature, "Proactivity," would enable the agent to create follow-up tasks without a new user prompt, according to Gizmodo.
Gizmodo reported that features surfacing first in Codex's command-line tool have tended to reach other product surfaces over time, and characterized a desktop Codex release as a possible future destination. OpenAI did not respond to Gizmodo's request for comment before publication.
What persistence changes
A persistent coding agent differs materially from a one-shot code-generation workflow. Rather than ending after a single response or task run, an agent with continuing execution can repeatedly inspect a repository, run tests, revise code, and potentially queue additional work. The reported reasoning-effort control is especially relevant because extended agent runs can consume substantial inference capacity.
For engineering teams, comparable always-on agent designs make execution boundaries central operational questions: which repositories, terminals, credentials, network resources, and deployment environments an agent can access, and when a human approval gate interrupts it. Autonomous follow-up task creation can also compound an initial error if task selection, permission controls, or evaluation signals are weak.
The available reporting does not establish a public launch date, supported environments beyond the command-line code discovery, or the final behavior of either Persistent mode or Proactivity. It also does not establish whether the features will ship in their currently reported form.
Key Points
- 1Wired identified code for a persistent Codex mode, potentially extending coding-agent work beyond a single prompted session.
- 2Reported reasoning-effort controls could give users a direct way to constrain compute consumption during extended autonomous runs.
- 3Comparable persistent-agent systems increase the importance of permissions, approval gates, observability, and reliable task-evaluation loops.
Scoring Rationale
A persistent execution mode for Codex would be a significant change to the product, although it is currently based on code discovery rather than a confirmed launch. The report matters to software practitioners because persistent agents raise practical compute-governance, access-control, and evaluation concerns.
Sources
Public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
