Thibault Sottiaux describes the route from research-oriented autonomous software engineering to Codex, including a Rust core shared across product interfaces. Gergely Orosz probes the engineering choices behind correctness, security and iteration speed, rather than treating an agent as merely a chat interface.
Thibault Sottiaux discusses open-source development as both a source of feedback and a coordination cost. Contributions can improve onboarding and reveal user needs, but low-quality submissions and competitors copying unfinished features also consume attention. Model choice and local versus cloud execution introduce further product and isolation tradeoffs.
Thibault Sottiaux frames the harness as support that anticipates what a model cannot yet do reliably. Research and engineering co-design the surrounding tools and instructions, then reconsider those scaffolds as models improve. These are the guest's account of Codex development, not independently verified capability benchmarks.
Thibault Sottiaux argues that automated review can increasingly handle correctness and security checks while people discuss intent, architecture and the guarantees a component must preserve. Gergely Orosz asks what remains valuable in human review and maintenance; clear interfaces and invariants become more important as more agents contribute concurrently.
Thibault Sottiaux describes the integration of local-agent capabilities into a managed cloud product and his own use of agents for research, reports and prototypes. His closing advice emphasizes curiosity, understanding unfamiliar systems and staying close to the community whose problems an engineer is solving.
Watch on YouTube
