AI + a16z How OpenAI Built Its Coding Agent
135 snips
Aug 29, 2025 
Alexander Embiricos, a product lead for Codex at OpenAI, shares thrilling insights into the development of AI coding agents. He discusses how Codex achieved astonishing success rates on GitHub in just its first month. The conversation dives into safety challenges of autonomous agents and the balance between coding automation and human oversight. Embiricos also highlights the transformative role of AI tools in education, where students collaborate with AI to gain real-world coding experience. It's a future-forward look at coding that blends innovation with human intuition.
AI Snips
Chapters
Transcript
Episode notes
Agents Are Reasoning Models With Tools
- OpenAI designed Codex as a reasoning model given tools and an environment to act like an agentic teammate.
- The product form factor chosen was a cloud agent with controlled permissions to explore safety and parallel work.
Late-Stage PRs Improve Merge Rates
- Codex's high PR merge rate reflects its cloud-agent form factor that works privately and completes work before opening PRs.
- That hidden pipeline increases success rates compared with agents that open PRs earlier.
Gate Network Access Before Merging
- Avoid allowing agents to run unvetted networked code without human review to reduce exfiltration risk.
- Gate network access and require explicit human confirmation before publishing or uploading artifacts.