OpenAI Tests Always-On Mode for Its Codex Agent
OpenAI is experimenting with a new 'Persistent mode' for its Codex AI agent, which would allow it to work continuously on tasks until explicitly halted. The feature is visible in the codebase but not yet publicly launched. The company says it is testing the capability with no immediate plans for release.
The Persistent mode setting appears within Codex's "reasoning effort" menu, where users currently choose how much computing power and time an AI model may use before responding. Code in OpenAI's public repository indicates that in this mode, the agent would keep working until explicitly deactivated, unlike existing modes that stop after minutes or hours even when tasks remain unfinished.
A related "proactivity" feature would instruct the agent to generate its own follow-up tasks after completing a user's request, drawing on past interactions and user knowledge across sessions. The agent could message users unprompted, though sparingly, and would require approval before altering anything outside the user's own system. OpenAI's head of core products described the repository as a "shared playground" where many ideas are explored.
An always-on AI agent could meaningfully shift how people delegate daily work, potentially automating ongoing responsibilities like scheduling, research, or administrative follow-through without constant human prompting. This may benefit busy professionals and less technical users, but could also raise concerns about autonomy, oversight, and unintended actions taken without explicit approval. The safeguards described suggest OpenAI is aware of these risks, yet the broader societal effect will depend on how reliably such agents respect boundaries and how much trust users place in them over time.