Dev Interrupted Scaffolding is coping not scaling, and other lessons from Codex | OpenAI’s Thibault Sottiaux
13 snips
Jan 27, 2026 
Thibault Sottiaux, an OpenAI Codex engineer who builds agentic coding systems, explains why removing scaffolding leads to true agent autonomy. He discusses agent-first design, simple scalable primitives over clever harnesses, open-sourcing Codex tooling, multi-agent scaling, portability trade-offs, and how agents reshape developer workflows and careers.
AI Snips
Chapters
Transcript
Episode notes
Agent-First Over Product-First
- Building an agent-first system reveals many unexpected, valuable use cases beyond a single product interface.
- Thibault Sottiaux says focusing on a general agent lets you later decide the best product form factor.
Vertical Integration Unlocks Capability
- Vertical integration lets research and engineering influence each other and enables trade-offs you can't make otherwise.
- Sottiaux argues coupling model and harness yields capability lifts you can't get with separate stacks.
Simplicity Prevents Capability Overhang
- Simplicity in the harness reduces constraints that create capability overhangs as models improve.
- Sottiaux recommends choosing primitives that scale with model capabilities to keep the system extensible.