Descriptions:
This 43-minute interview hosted by Peter Yang features Alex and Romain, members of OpenAI’s Codex team, demonstrating how they actually build new Codex features using Codex itself — a rare first-party look at the tool’s real-world development workflow. The conversation covers the full stack: GPT 5.4 for complex, multi-file tasks requiring deep reasoning; Codex Spark for rapid iteration at roughly 1,200 tokens per second; and the Codex app’s design philosophy of separating agent work from any single local folder.
Alex and Romain explain that the Codex team writes almost no formal specs — typically 10 bullet points or fewer — and that designers on the team now write more code than engineers did six months ago. They discuss the architectural decision to build a standalone Codex app alongside the IDE extension (available for VS Code, Cursor, and Windsurf), sharing that both tools run on the same open-source Rust-based Codex harness under the hood. The interview also surfaces an honest account of internal debate around whether to build the app at all versus doubling down on the IDE extension.
For developers and product teams thinking about how to structure agentic coding workflows, the conversation offers concrete, experience-backed perspectives on delegating to cloud agents, handling partial task completion, and working with multiple parallel agents — directly from the people building and daily-driving the tooling.
📺 Source: Peter Yang · Published April 05, 2026
🏷️ Format: Interview
