Krux
May 10, 2026
OpenAI's Codex Now Runs Autonomously For 7 Hours Straight
Published: May 10, 2026 at 12:13 AM
Updated: May 10, 2026 at 12:13 AM
100-word summary
OpenAI just published the playbook for letting AI agents write code unsupervised. Their GPT-5-Codex system now runs autonomously for over seven hours on complex tasks, guided by three-tier approvals and an auto-review subagent that decides which changes need human eyes. Network access stays disabled by default, and everything runs in a sandbox. The striking part: container caching cut median task completion by 90%, making follow-up requests nearly instant. OpenAI tested malware refusal and prompt-injection resistance, both scoring above 0.97. The real shift is turning coding assistants into coding colleagues that handle routine changes overnight while you sleep.
What happened
OpenAI just published the playbook for letting AI agents write code unsupervised. Their GPT-5-Codex system now runs autonomously for over seven hours on complex tasks, guided by three-tier approvals and an auto-review subagent that decides which changes need human eyes. Network access stays disabled by default, and everything runs in a sandbox. The striking part: container caching cut median task completion by 90%, making follow-up requests nearly instant. OpenAI tested malware refusal and prompt-injection resistance, both scoring above 0.97.
Why it matters
The real shift is turning coding assistants into coding colleagues that handle routine changes overnight while you sleep.