OpenAI is building agents that don't clock out when you close the browser tab.
The Summary
- OpenAI is developing a "persistent" mode for its Codex agent that keeps working after you walk away, until explicitly told to stop
- This isn't about better chatbots — it's about agents that operate on their own timeline, not yours
- The shift from "assistant you summon" to "employee that keeps working" changes what AI companies are actually selling
The Signal
Code reviewed by Wired shows OpenAI building infrastructure for agents that stay awake. The feature lets Codex — their code-writing model — continue executing tasks proactively until a user explicitly puts it to sleep. Not "pause until next prompt." Not "wait in the background." Actually working. On its own. While you're in a meeting or asleep or living your life.
This is the quiet crossing of a major threshold. Every AI tool you use today operates on request-response. You ask, it answers. You prompt, it generates. The interaction ends when you stop talking to it. Persistent agents flip that model. They operate on their own loop.
"The shift from summoned assistant to autonomous employee changes what AI companies are selling."
The implications stack quickly:
- Compute costs move from per-query to per-hour (or per-day). Right now, OpenAI charges you when you ping GPT-4. Persistent agents rack up compute time whether you're watching or not. That's a different revenue model and a different infrastructure challenge.
- Reliability becomes existential. If an agent is running unsupervised for six hours, a hallucination isn't just wrong — it's six hours of compounding mistakes. Error correction moves from "verify the output" to "verify the process."
- The product surface expands beyond chat. Persistent agents need dashboards, activity logs, stop buttons, sandboxed environments. You're not building a chatbot anymore. You're building a control panel for autonomous workers.
Codex writing code while you sleep sounds like productivity porn. But the harder question is trust. Not "can it write the function" but "should it be allowed to keep iterating on the codebase for three hours without human checkpoints?" OpenAI hasn't published safety guidelines for this yet. They're building the capability first.
The Implication
If persistent agents work, the next wave isn't better prompts — it's better management systems. Companies will need tooling to monitor, pause, and roll back agent activity. Expect a Cambrian explosion of "agent ops" platforms: logging, version control, audit trails, kill switches. The picks and shovels for people running workforces that don't have lunch breaks.
Watch for pricing changes. OpenAI's current API model doesn't map to "agent runs for 8 hours straight." They'll need subscriptions, compute pools, or some hybrid. How they price persistence will determine whether this is a tool for enterprises or individuals.