How OpenAI’s Codex Team Uses Their Coding Agent

Rhea Purohit interviews Thibault Sottiaux and Andrew Ambrosino from OpenAI’s Codex team about their internal workflows and product strategy. The conversation covers how the team that builds the coding agent actually uses it day-to-day, the workflows they rely on most heavily, and a counterintuitive finding: as model speed increases, new bottlenecks emerge in the human review and integration steps.

RDCO mapping: The speed-creates-new-bottlenecks insight matches our experience. When the agent finishes faster, the constraint shifts to our ability to review and validate output. Worth considering how our autonomous loop handles this — perhaps by building better self-verification into the agent before surfacing results.