OpenAI Codex Lead on Building Sora Android in 18 Days

How Alexander Embiricos Sees the Future of Coding Agents

This is Alexander Embiricos - product lead for OpenAI Codex - explaining why their coding agent grew 20x since August and why the vision is much bigger than autocomplete.

"A really smart intern that refuses to read Slack." This is how Alexander describes Codex today. It can do amazing things when you give it proper guidance, but it doesn't check DataDog unless you ask. It doesn't participate in planning. It doesn't know what happened in the standup. That's what they're building toward: a teammate, not a tool.

The Sora Android app was built in 18 days. A fully new app, then 10 days later (28 days total) it went public. This is Codex accelerating OpenAI internally. Codex is also on-call for its own training runs - writing infrastructure code, code review catching configuration mistakes.

Compaction enables 24-hour agent runs. Models routinely work overnight now. When they approach context limits, a new capability called "compaction" lets them prepare a compressed context, restart in a fresh window, and continue. This required coordinated work across model, API, and harness layers.

The cloud product was too far in the future. Codex cloud launched first - a remote computer you could delegate to massively in parallel. Great for OpenAI employees used to reasoning model prompting. Not great for general adoption. The unlock: land with users in the IDE and CLI first, then graduate them to delegation over time.

"Would I have written this prompt? Maybe 50/50." That's good enough. The agent's job isn't to be perfect - it's to maintain consistency and run 24/7. Karpathy's use case: give it your gnarliest bugs, let it run for an hour. It solves problems that nothing else can.

10 Insights From OpenAI Codex's Product Lead

• 20x growth since August - Codex models most served in API
• Sora Android app in 18 days - Hit #1 in app store at 28 days
• Codex on-call for own training - Writes infrastructure, reviews code
• Compaction - Enables 24-hour runs by compressing context
• GPT 5.11 Codex Max - 30% faster at same tasks, smarter at higher reasoning
• Cloud product too futuristic - IDE/CLI first, then graduate to delegation
• Proactivity is the mission - Thousands of helpful moments per day, not tens
• "Gnarliest bugs" use case - Karpathy gives hardest problems to Codex overnight
• 50/50 prompts are fine - Consistency and 24/7 availability beat perfection
• Dogfooding signal differs - OpenAI employees too comfortable with reasoning prompts

What This Means for AI-Powered Development

Codex isn't trying to be better autocomplete - it's trying to be a teammate that works while you sleep. The shift from tool to colleague requires proactivity: thousands of helpful moments daily, not tens. When your coding agent is on-call for its own training runs, the recursive loop of AI improving AI has begun.