Summary
The Pragmatic Engineer’s Gergely Orosz sits down with Tibo Sottiaux, who led OpenAI’s Codex team from its inception and oversaw its eventual merge into ChatGPT. The conversation covers the full arc of Codex’s development: why the team chose to build the agent harness in Rust despite the model having limited Rust training data at the time, how Codex quickly became the default internal tool for OpenAI engineers themselves (‘Have you asked Codex?’ is apparently the first answer new hires get to any question), and the architectural decision to run all tool executions inside a sandbox by default, prompting users only when elevated permissions are required.
Sottiaux speaks candidly about how AI is reshaping software engineering workflows — not just code generation but code review, where the reviewer’s role is shifting toward intent verification rather than line-by-line scrutiny. He also discusses the economics of maintenance: tasks that once required ongoing human upkeep are increasingly becoming candidates for full automation, which changes the calculus of whether to build or buy entirely. The episode includes a Fortune 10 company anecdote (also referenced elsewhere in the AI space) about engineers hitting $100/day Claude Code rate limits and structuring their afternoons around the 4 p.m. reset — illustrating how deeply AI tooling is now embedded in professional engineering workflows.
📺 Source: The Pragmatic Engineer · Published September 09, 2026
🏷️ Format: Podcast
