The Stack Overflow Podcast Dogfood so nutritious it’s building the future of SDLCs
13 snips
Feb 24, 2026 
Thibault Sottiaux, engineering lead on Codex at OpenAI, builds agentic coding tools and secure SDLC tooling. He explains how agentic systems differ from chat assistants. He discusses dogfooding Codex inside OpenAI, sandboxing and safety, automated code reviews, onboarding agents to large codebases, multi-agent refactors, and ambitions for proactive, memory-enabled SDLC automation.
AI Snips
Chapters
Transcript
Episode notes
Always Start Agents In A Sandbox
- Run coding agents in a sandbox by default and restrict network and filesystem access.
- Thibault recommends configuring outside the sandbox only when you understand the risks and supervise closely.
Codex Dogfoods Itself For PR Reviews
- The Codex team dogfoods Codex heavily and measures compute deployed per individual as a productivity metric.
- Codex now automatically reviews 100% of PRs at OpenAI, creating an ambient intelligence safety net.
Treat AI Review As A Safety Net Not Approval
- An agent approving a PR is insufficient; its review is a safety net, not a merge signal.
- Thibault warns against the pitfall of Codex writing code and then stamping its own PR without human oversight.