Tibo Sottiaux leads the Codex team at OpenAI; this episode covers his path from applied mathematics and supply-chain optimization through Google and DeepMind to building Codex, the open-source, Rust-based coding agent that now runs both locally and in ChatGPT, and how AI is reshaping software development practices at OpenAI.
Tibo’s path to OpenAI
Grew up in a tiny Belgian village; got into computers early as his only window to the world.
Studied applied mathematics, ran a startup optimizing pharmaceutical clinical-trial supply chains (Monte Carlo, stochastic multi-stage optimization), then joined Google in 2015.
At Google: first on a “make the web faster on mobile” project (cancelled after two years — taught him to always question impact and product-market fit), then Google Maps reviews, then DeepMind research infrastructure.
At DeepMind: part of a small group that built an internal chatbot a year before ChatGPT using early LLMs; it spread virally inside DeepMind but Google’s product-launch machinery made external release nearly impossible.
Joined OpenAI in 2024 because he wanted a mission-driven culture where research and product co-design, a tight feedback loop with users, and a small, autonomous team (ChatGPT had ~20 engineers at the time).
The early days of Codex
Started as an internal research-infrastructure effort: train models on OpenAI’s Python codebase to help researchers code faster.
Merged with the “ASWE” (Autonomous Software Engineer) effort; Greg Brockman pushed to make it a product for the world, not just OpenAI.
First launch: cloud Codex (high friction, no PMF), then Codex CLI; continuous push to let models do more of the work.
Why Codex was built in Rust
Deliberate separation between the agent core (Rust) and the product interface (TypeScript) to enforce clean boundaries and prevent intertwining.
Rust gives compile-time correctness, security, efficiency, and scales well; internal models were already decent at Rust.
Turned out Rust is also a good fit for agents because static verification catches errors early.
Could have succeeded in TypeScript/Python and rewritten later, but the Rust boundary forced architectural discipline from day one.
Why Codex is open source
A coding agent should be able to improve itself; open source enables a community of contributors using the agent to improve the agent.
Early stage: no one has all the answers; building in the open invites tinkering, exploration, and learning from the broader community.
Benefits: instant onboarding (new hires already know the repo), good external contributions, team energy from visible community engagement.
Downsides: artificial repo boundaries vs. internal monorepo, competitors copying features before release, flood of low-quality contributions requiring triage.
Codex plays nice with other models
Coupling the harness to OpenAI models would feel like lock-in; users would just fork and add 10 lines for another provider.
Optionality benefits users (try new models without changing setup) and OpenAI (gets feedback on what works elsewhere).
Enterprise customers demand model optionality; winning on merit (best model + best harness) attracts better talent than winning on lock-in.
How the harness works today
Default: runs sandboxed on the user’s local machine; asks permission for commands needing outside-sandbox privileges.
Optional: “cloud mode” runs the same harness in a managed Kata-container VM (same as ChatGPT’s code interpreter), streaming only I/O to the laptop.
Future: seamless hybrid — partial local, partial cloud — because models will leverage more compute than laptops can provide.
Cloud dev environments never took off due to setup/maintenance cost; agents now make that setup nearly free (agent configures the remote box).
Harness and model improvements
Harness is “a step ahead of the model”: provides guardrails, developer messages, crutches so the model behaves reliably today.
Over time, model training absorbs those behaviors (e.g., running tests automatically), so the harness and developer messages shrink.
Co-design process: research and engineering jointly decide whether a gap is a harness fix (fast) or a model fix (months); agents analyze feedback across all coding domains to prioritize.
North star: a delightful, simple, personal AGI that knows you, has access to resources, can take risky actions with verification, controllable via natural language/voice.
The SDLC behind Codex
Onboarding: “Have you asked Codex?” — Codex is plugged into Slack, docs, code; new hires use it to understand projects, decisions, people.
Work in public channels, broad doc permissions so agents can reason across context.
Principles over process: care about user, product coherence, model trajectory; avoid building 10k-line crutches for model flaws.
Shipping: PRs can go to millions of users same/next day; ownership and care replace heavy gates; automation handles reviews, deploys, regression detection.
Code reviews at Codex
Built superhuman code-review models (logic, security, deep dependency reasoning); now baked into mainline models.
Security reviews mandatory and automated: PRs blocked on critical vulnerabilities.
Human review shifting from correctness/security (automated) to intent discussion: “What are you trying to do? Is that the right thing?”
Future: agree on the “box” (invariants, resource limits, data access); internals of the box can change freely without discussion.
Maintenance and architecture
Maintenance = tax to keep things running; large parts becoming automated (dependency upgrades, security patches via changelog reasoning).
Re-architecture cost plummeting: what took years now happens in weekends; good abstractions (“boxes with invariants”) still matter to enable fast iteration.
Models increasingly capable of long-term architectural thinking (modularity, future extensibility).
Software lifecycle compresses: 100 agents can contribute in a weekend; engineers must design for that velocity from the start.
How AI tools expand what engineers can do
Not about losing craft; about staying in flow and solving more ambitious problems.
Benchmarking, trade-off analysis, prototyping now take seconds; engineers become better by focusing on outcomes.
OpenAI uses this to run inference more efficiently, deploy more effective compute.
No one at OpenAI has said “this isn’t fun” — the problem space (scientific breakthroughs, human-centric AGI) is far from exhausted.
Personal workflow: dictates tasks to ChatGPT/Codex on mobile between meetings; gets reports, slide decks, code explorations in minutes; builds prototypes on weekends to “flush ideas out of the system.”
The Merge: ChatGPT + Codex
Major engineering challenge: merging fully local Codex stack with ChatGPT’s managed cloud stack, efficiently enough for the $20 Plus plan.
Result: Codex harness runs in a powerful cloud VM (internet, Blender, model training) accessible via ChatGPT “Work” toggle.
Unified system goal: same capabilities whether in Codex or ChatGPT; temporary toggle state.
Codex itself acted as project journalist — ingested all Slack, docs, debates, decisions — creating a full recounting of the merge process.
How Tibo uses Codex and ChatGPT
Calendar is “Tetris” but cheerful because agents multiply throughput.
Mobile-first: dictates tasks/questions to ChatGPT Work (custom skills/instructions for report style); gets answers in 30 min.
Uses it for everything: public sentiment, production logs, deprecation lists, team status, code explorations, weekend prototypes.
Long-running tasks: /goal command (harness crutch) kept model on track for days; new models need it less — can just say “work for a week.”
Advice for engineers who want to work in AI
Deep curiosity + ability to grok systems quickly; agents accelerate this but the drive to ask “why” five levels down is core.
Be in tune with the community you’re building for: crisp taste, clear intent, ability to explain what you’re achieving and for whom.
Fundamentals still matter: clarity of thought, architectural judgment, understanding invariants — these become more valuable, not less, as implementation gets cheaper.
