OpenAI is teaching its coding agent to keep working long after you stop watching it.
Internal code under development shows OpenAI building a feature for Codex, its coding agent, that lets it continue working on tasks proactively instead of waiting for a new prompt each time. The feature would keep the agent active until a user explicitly puts it 'to sleep,' the company's own term for pausing it. The code is still in development, not a shipped feature, and no release timeline has surfaced. It is not yet clear what 'working proactively' actually covers, whether that means running tests, fixing bugs, or opening pull requests without a human asking first.
This is a bigger shift than a new settings toggle. It moves Codex from a tool that waits for instructions to one whose default state is 'on,' putting the burden on the user to actively shut it off. That raises real questions about compute costs, unsupervised changes to a codebase, and who signs off on work nobody explicitly requested in the moment.
Every major coding-agent vendor is chasing more autonomy this year. Whether 'keeps working until told to stop' turns out to be genuinely useful independence, or just a longer meter running on your API bill, is the part still worth watching.