Photo: Merlin Lightpainting / Pexels
Code reviewed by Wired reveals a feature that lets Codex keep coding autonomously until it's explicitly told to stop, marking a shift from assistant to independent software engineer.
OpenAI is building a version of Codex that doesn’t clock out. Code reviewed by Wired reveals the company is developing a persistent agent feature that enables its AI coding tool to continue working proactively until it is “put to sleep,” a quiet but meaningful leap from chatbot-style coding help to something closer to an always-on software engineer.
Instead of responding to prompts and waiting for the next one, Codex would autonomously schedule future work, wake itself up, and continue grinding on long-term projects spanning days or weeks without human intervention.
From assistant to autonomous agent #
One of the key mechanisms making this possible is what OpenAI calls “thread automations.” Think of them as alarm clocks for an AI agent. They preserve context for ongoing tasks while enabling background checks or follow-ups, so the agent doesn’t lose its train of thought between sessions.
There’s also a “/goal” mode that lets users define specific milestones. Once a goal is set, Codex works persistently toward achieving it across extended timelines without losing progress.
Internal tests reportedly showed agents could run for approximately 25 hours on complex tasks, generating extensive code and iterating through workflow loops.
The infrastructure play #
Making an AI agent that runs continuously requires more than clever prompting. It requires serious cloud infrastructure, which explains OpenAI’s acquisition of Ona in June 2026.
The Ona deal was designed to create persistent execution environments that let Codex operate within customer-controlled infrastructure, even when the user’s own devices are offline.
What this means for the developer economy #
The risks are real, though. An agent that runs autonomously for 25 hours can generate a lot of bad code just as easily as good code. Quality control, review processes, and rollback mechanisms become critical when the AI doesn’t for approval between steps. The “put to sleep” framing in Wired’s reporting suggests OpenAI is aware that the default state of these agents is “on,” which introduces failure modes that don’t exist in prompt-response systems.
The adoption curve for agent-optimized models has been steep since early 2025, reflecting strong developer interest in tools that go beyond suggestion and into execution.
Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our