AI Radar

Noteworthy AI conversations and writing, summarized from publisher notes.

How OpenAI Built Its Coding Agent

· The a16z Show

Alexander Embiricos, who leads product for Codex at OpenAI, walks through how the new cloud-based Codex coding agent was conceived and built. The core insight: an agent is a reasoning model given access to tools (like a terminal) and a specific environment, tuned for a particular task. The team had an early internal breakthrough watching a reasoning model edit React code via "sight reading" the source — and they decided the right form factor was a cloud agent working on its own computer, not an in-IDE copilot. After 35 days in public research preview, Codex has opened roughly 400K PRs with an ~80% merge rate, far ahead of other coding agents. Embiricos argues this is partly a reflection of the form factor — Codex does extensive work in an isolated environment and shows its diff before opening a PR, which doubles as a safety mechanism against prompt injection. He describes prompt injection as a layered problem (the prompt itself, mid-task actions, and final outcomes like exfiltration) requiring defenses at every layer. Usage surprises: building new features dominates over debugging, and multi-turn iteration turned out far more popular than internal OpenAI users expected — exposed by a real bug in the three-turn flow. He frames the product like a "slot machine" — best-of-N generation, where the human (or eventually an evaluator model) curates outputs, much like Midjourney's grid. On the roadmap: faster container cold starts, deeper integration with IDEs/issue trackers, support for air-gapped and on-prem enterprise deployments, and a long-term vision of Codex as an ambient "teammate" that picks up work from team workspaces rather than waiting for prompts.

AI Radar wrote this summary from the publisher’s notes with AI, not from a transcript. How it works.